GitHub - crowd-dataset/crowd: This repository contains code that extracts YouTube videos based on a mapping.csv file and performs object detection using YOLOv11. The primary objective of this work is to evaluate pedestrian behaviour in a cross-country or cross-cultural context using freely available YouTube videos.

City Road Observations With Dashcams (CROWD)

This repository contains code that extracts YouTube videos based on a mapping.csv file and performs object detection using YOLOv11. The primary objective of this work is to evaluate pedestrian behaviour in a cross-country or cross-cultural context using freely available YouTube videos.

This study presents a comprehensive cross-cultural evaluation of pedestrian behaviour during road crossings, examining variations between developed and developing states worldwide. As urban landscapes evolve and autonomous vehicles (AVs) become integral to future transportation, understanding pedestrian behaviour becomes paramount for ensuring safe interactions between humans and AVs. Through an extensive review of global pedestrian studies, we analyse key factors influencing crossing behaviour, such as cultural norms, socioeconomic factors, infrastructure development, and regulatory frameworks. Our findings reveal distinct patterns in pedestrian conduct across different regions. Developed states generally exhibit more structured and rule-oriented crossing behaviours, influenced by established traffic regulations and advanced infrastructure. In contrast, developing states often witness a higher degree of informal and adaptive behaviour due to limited infrastructure and diverse cultural influences. These insights underscore the necessity for AVs to adapt to diverse pedestrian behaviour on a global scale, emphasising the importance of incorporating cultural nuances into AV programming and decision-making algorithms. As the integration of AVs into urban environments accelerates, this research contributes valuable insights for enhancing the safety and efficiency of autonomous transportation systems. By recognising and accommodating diverse pedestrian behaviours, AVs can navigate complex and dynamic urban settings, ensuring a harmonious coexistence with human road users across the globe.

The dataset is available on kaggle. The dataset shall soon be made available on a permanent FAIR storage.

Dataset overview

Footage: 36,204.3 hours (1,508.5 days)
Videos: 82,484 unique videos split into 107,407 segments
Localities (cities): 10,538
Countries and territories: 238
Continents: 6
YOLO detections: 91,983 processed segments (one detection CSV each)
Detected objects (unique tracked objects, confidence ≥ 0.0): 34,582,015 persons, 97,105,972 cars, 2,544,466 bicycles, 11,293,405 motorcycles, 3,571,231 buses, 14,385,922 trucks
Last updated: 2026-11-02 00:25 UTC

All 238 countries and territories

#	Country	Cities	Videos	Footage (h)	Analysed (h)	Persons	Cars	Bicycles	Motorcycles	Buses	Trucks
1	🇺🇸 United States	2,359	21,326	10,511.1	8,659.4	4,820,542	29,799,823	294,021	156,255	510,360	4,296,528
2	🇦🇺 Australia	213	4,124	1,729.0	1,380.7	720,050	5,101,216	48,930	52,316	109,753	770,771
3	🇬🇧 United Kingdom	379	3,238	1,703.9	1,510.1	2,867,090	3,590,817	261,883	154,434	262,716	570,646
4	🇷🇺 Russia	212	3,424	1,522.9	1,447.9	844,437	4,317,947	17,210	7,814	168,343	513,024
5	🇵🇭 Philippines	414	3,688	1,350.0	1,316.8	2,427,410	2,407,750	72,700	1,487,797	310,275	939,564
6	🇨🇦 Canada	239	3,733	1,273.7	1,028.5	818,846	3,272,258	71,926	20,308	57,635	395,753
7	🇻🇳 Vietnam	79	3,074	1,196.8	1,115.7	3,060,351	2,014,083	90,216	3,001,694	126,115	301,648
8	🇮🇩 Indonesia	177	3,192	1,033.6	1,018.6	2,135,172	2,062,675	26,720	2,098,743	113,916	501,032
9	🇰🇷 South Korea	84	855	822.8	818.9	386,095	1,963,479	21,093	56,226	113,188	188,634
10	🇫🇷 France	369	1,911	750.4	717.4	869,861	1,906,387	101,869	124,879	47,050	197,996
11	🇨🇳 China	160	1,275	714.3	629.5	506,472	1,961,536	50,851	197,943	103,161	211,319
12	🇩🇪 Germany	275	1,486	667.6	607.4	555,786	1,888,491	186,722	35,261	59,918	208,708
13	🇯🇵 Japan	241	1,802	623.1	595.7	439,838	1,275,282	54,810	24,860	86,925	314,918
14	🇧🇷 Brazil	131	1,472	613.6	540.4	551,015	1,703,926	34,597	155,795	55,591	198,772
15	🇳🇱 Netherlands	232	1,474	556.5	437.2	517,322	1,094,479	353,908	55,627	32,181	116,122
16	🇮🇳 India	303	1,531	488.3	476.8	1,358,956	1,161,598	66,826	757,324	134,181	363,542
17	🇺🇦 Ukraine	97	1,144	458.7	436.7	346,811	1,566,587	9,078	4,834	71,397	212,145
18	🇹🇭 Thailand	50	818	401.4	376.3	485,346	1,069,656	8,685	365,006	50,806	228,714
19	🇲🇾 Malaysia	70	652	372.2	282.2	142,606	893,879	1,759	85,354	17,961	78,616
20	🇲🇲 Myanmar	19	536	371.5	371.5	393,614	1,646,979	44,384	42,070	101,213	260,286
21	🇦🇪 United Arab Emirates	10	905	370.4	306.1	105,575	885,948	4,006	15,784	42,977	113,247
22	🇵🇱 Poland	283	918	361.9	337.0	231,202	1,002,233	21,276	7,242	33,761	87,286
23	🇮🇷 Iran	109	1,071	334.3	264.1	177,639	1,444,918	3,631	36,031	14,546	114,312
24	🇸🇪 Sweden	174	1,036	284.9	266.5	141,949	482,271	22,144	2,821	18,749	51,592
25	🇹🇷 Türkiye	88	572	272.4	226.7	285,629	942,601	4,225	36,528	56,256	133,176
26	🇮🇹 Italy	175	691	267.0	261.3	282,142	908,873	23,915	101,935	17,661	69,447
27	🇩🇰 Denmark	22	570	257.7	250.7	390,428	780,277	200,153	8,317	20,282	61,102
28	🇦🇷 Argentina	63	411	257.1	256.5	367,849	931,207	35,207	43,315	30,270	101,251
29	🇷🇴 Romania	99	679	250.5	240.6	197,719	1,221,247	6,691	5,080	27,652	99,364
30	🇫🇮 Finland	44	387	239.9	235.5	134,640	495,813	14,687	2,445	21,620	55,653
31	🇪🇸 Spain	253	664	233.2	225.5	291,818	649,738	15,891	65,026	23,092	65,410
32	🇨🇿 Czechia	114	477	230.1	214.6	222,975	584,974	5,613	5,080	23,244	55,781
33	🇰🇭 Cambodia	7	427	222.3	221.8	606,066	561,842	13,545	623,624	38,992	181,902
34	🇦🇹 Austria	49	907	211.2	201.6	133,417	563,773	18,117	14,410	23,235	84,480
35	🇺🇬 Uganda	19	522	195.0	84.9	501,747	190,827	15,869	251,438	17,576	64,307
36	🇭🇺 Hungary	35	565	190.6	171.0	311,642	1,114,936	14,911	8,344	51,006	125,623
37	🇧🇬 Bulgaria	100	736	176.5	162.6	85,581	569,432	1,797	1,204	18,114	48,875
38	🇵🇰 Pakistan	39	426	167.7	146.7	490,898	509,985	16,387	250,458	28,994	123,420
39	🇸🇬 Singapore	1	417	163.8	152.6	84,721	384,215	5,235	20,853	33,904	61,744
40	🇳🇿 New Zealand	162	620	163.4	144.9	81,718	433,001	4,647	3,418	13,862	56,781
41	🇻🇪 Venezuela	28	659	162.7	147.7	118,784	408,059	5,065	41,064	9,027	79,602
42	🇮🇪 Ireland	65	561	158.3	138.4	114,472	342,626	10,742	2,226	14,161	41,970
43	🇲🇽 Mexico	73	313	146.2	138.7	143,106	493,212	6,527	23,581	15,433	72,687
44	🇳🇴 Norway	68	431	143.0	137.7	89,964	311,168	7,494	2,000	13,929	38,717
45	🇿🇦 South Africa	59	490	142.2	130.9	95,473	361,609	1,715	4,681	13,247	61,876
46	🇸🇾 Syria	13	353	139.4	124.3	230,775	568,753	6,624	17,705	17,508	86,472
47	🇧🇾 Belarus	83	281	139.2	81.4	47,143	204,755	1,607	417	10,628	27,332
48	🇭🇰 Hong Kong	11	230	107.8	59.1	108,426	106,587	1,444	2,993	19,606	19,712
49	🇪🇬 Egypt	21	272	107.7	53.4	78,141	268,992	737	8,549	12,471	33,615
50	🇧🇩 Bangladesh	28	297	107.2	107.2	426,287	117,440	75,686	73,689	41,201	87,546
51	🇵🇹 Portugal	75	231	105.3	104.4	91,994	322,672	2,144	7,495	7,753	32,103
52	🇬🇲 Gambia	35	394	102.9	95.4	246,297	318,516	14,756	21,032	8,768	66,261
53	🇨🇭 Switzerland	96	309	101.9	92.5	109,911	204,142	21,512	13,549	9,618	24,486
54	🇯🇴 Jordan	16	330	98.3	78.3	47,950	357,720	181	983	7,339	40,130
55	🇳🇬 Nigeria	35	239	96.2	66.3	263,110	220,926	1,860	33,764	18,502	56,569
56	🇸🇱 Sierra Leone	47	185	79.5	79.1	187,292	51,000	1,879	51,201	4,608	23,656
57	🇨🇱 Chile	43	161	72.0	26.4	27,553	83,092	969	1,381	3,246	15,475
58	🇲🇳 Mongolia	4	130	71.4	71.4	91,858	344,610	1,883	1,686	9,246	26,363
59	🇭🇷 Croatia	89	282	71.3	71.2	53,717	253,352	4,602	4,328	5,525	19,844
60	🇨🇾 Cyprus	19	256	67.1	66.4	27,722	228,547	1,082	5,023	4,182	28,797
61	🇬🇷 Greece	40	196	66.9	66.5	77,114	293,048	2,854	26,381	6,361	26,228
62	🇪🇪 Estonia	8	312	66.6	65.3	34,638	184,453	1,216	420	8,537	17,236
63	🇧🇪 Belgium	60	190	62.9	62.6	71,821	190,482	15,473	3,904	7,398	24,351
64	🇹🇼 Taiwan	39	186	62.6	43.7	51,993	127,694	2,546	46,958	4,547	12,812
65	🇧🇦 Bosnia and Herzegovina	61	296	62.4	62.0	69,752	222,699	1,056	702	5,842	16,960
66	🇨🇴 Colombia	36	131	61.7	50.9	117,370	129,247	6,718	51,817	7,850	21,360
67	🇺🇿 Uzbekistan	13	180	54.8	53.6	46,774	355,390	722	332	6,388	27,457
68	🇵🇪 Peru	29	95	53.6	53.1	88,967	163,413	5,474	13,113	9,663	31,275
69	🇦🇿 Azerbaijan	8	122	51.2	45.5	46,348	235,810	694	1,139	6,123	21,861
70	🇷🇸 Serbia	37	156	49.4	46.5	44,020	186,903	1,154	1,225	5,946	14,798
71	🇲🇦 Morocco	37	138	48.7	47.3	93,249	201,231	3,877	18,016	5,101	22,238
72	🇱🇰 Sri Lanka	17	133	48.6	47.3	74,781	106,111	1,066	25,811	17,807	41,272
73	🇸🇦 Saudi Arabia	11	122	46.2	32.4	5,162	169,306	118	248	2,843	18,349
74	🇶🇦 Qatar	5	98	44.3	44.3	10,813	125,797	434	1,284	5,347	16,648
75	🇪🇹 Ethiopia	8	77	44.1	42.2	123,021	125,171	542	1,984	9,835	34,724
76	🇬🇪 Georgia	12	119	42.8	42.4	33,673	201,005	246	1,328	7,185	22,399
77	🇱🇻 Latvia	20	124	38.4	38.0	26,293	105,683	1,706	748	4,275	12,473
78	🇳🇵 Nepal	4	120	38.2	38.2	225,002	72,918	6,060	131,456	10,910	20,116
79	🇱🇧 Lebanon	62	157	35.4	35.2	23,417	163,967	184	7,670	2,710	18,124
80	🇨🇺 Cuba	25	165	34.4	34.3	117,049	49,556	4,702	16,595	3,856	14,402
81	🇱🇹 Lithuania	11	86	33.9	33.5	32,986	106,385	2,629	716	3,324	7,526
82	🇩🇿 Algeria	8	62	33.8	14.5	26,027	68,925	133	1,304	1,372	6,901
83	🇵🇷 Puerto Rico	60	159	33.1	19.4	4,934	70,393	115	434	442	9,614
84	🇲🇼 Malawi	51	115	32.9	31.5	78,660	32,742	10,602	5,436	1,821	9,274
85	🇰🇿 Kazakhstan	26	116	31.9	30.8	22,288	134,602	260	202	4,090	13,041
86	🇦🇲 Armenia	10	105	30.8	23.9	21,166	153,527	126	319	4,473	15,760
87	🇰🇪 Kenya	58	118	27.7	19.9	70,040	50,504	1,168	12,731	4,621	13,321
88	🇪🇨 Ecuador	14	58	26.7	22.8	28,852	75,067	686	4,759	3,540	13,055
89	🇲🇰 North Macedonia	21	75	26.6	24.4	26,662	98,030	1,779	1,072	2,519	7,901
90	🇺🇾 Uruguay	23	90	25.3	25.3	33,471	99,796	2,762	5,329	1,627	12,799
91	🇲🇩 Moldova	14	111	25.3	24.2	26,690	113,963	302	217	5,094	12,595
92	🇱🇦 Laos	11	62	23.1	22.9	24,568	53,800	528	23,150	1,272	20,247
93	🇧🇴 Bolivia	20	55	21.5	21.5	33,175	52,027	775	5,899	3,121	15,024
94	🇲🇨 Monaco	1	45	20.3	20.3	38,846	32,963	1,033	13,199	1,598	4,440
95	🇮🇱 Israel	18	68	19.9	19.0	15,117	74,185	2,391	3,131	2,717	5,690
96	🇦🇱 Albania	23	63	15.5	14.8	19,340	52,118	1,616	1,220	1,398	5,443
97	🇽🇰 Kosovo	31	93	15.4	15.4	16,804	66,243	815	885	1,405	8,619
98	🇬🇹 Guatemala	9	23	15.3	15.3	18,049	46,886	685	9,016	1,774	12,444
99	🇲🇪 Montenegro	24	91	14.6	10.2	9,014	42,328	173	753	965	4,093
100	🇨🇬 Congo (Congo-Brazzaville)	5	38	13.9	11.8	29,968	24,408	229	3,092	1,062	5,312
101	🇮🇸 Iceland	14	40	13.2	12.8	7,533	31,458	475	68	955	4,582
102	🇱🇾 Libya	3	39	13.2	13.2	4,990	64,240	79	115	299	7,120
103	🇬🇭 Ghana	12	43	12.3	11.4	24,607	31,824	671	4,579	2,247	8,944
104	🇹🇲 Turkmenistan	1	85	11.9	11.9	3,824	26,509	47	6	817	2,832
105	🇮🇶 Iraq	10	41	11.6	9.8	4,343	38,747	120	480	539	4,969
106	🇴🇲 Oman	10	32	10.9	10.9	3,127	45,003	156	60	744	3,576
107	🇸🇻 El Salvador	6	13	10.8	10.8	18,428	34,086	379	4,727	1,585	7,774
108	🇭🇳 Honduras	8	23	10.3	10.3	13,436	34,316	572	5,064	1,410	7,641
109	🇬🇳 Guinea	7	45	10.1	6.3	20,346	18,081	147	6,727	454	3,761
110	🇹🇿 Tanzania	13	50	9.4	9.0	28,697	22,586	1,073	8,485	2,785	6,395
111	🇵🇦 Panama	7	22	9.4	9.4	6,520	36,612	151	679	720	6,998
112	🇬🇾 Guyana	6	20	9.0	9.0	9,187	28,212	748	864	1,186	7,344
113	🇸🇮 Slovenia	26	49	8.9	8.5	10,356	27,123	1,574	357	900	2,071
114	🇧🇭 Bahrain	11	42	8.8	8.8	2,042	28,027	182	229	750	3,663
115	🇯🇲 Jamaica	10	31	8.7	8.3	9,336	21,367	223	356	1,471	3,556
116	🇰🇬 Kyrgyzstan	7	40	8.4	8.4	6,778	33,286	161	72	1,501	4,983
117	🇵🇸 Palestine	10	31	8.3	8.3	5,509	29,388	95	245	812	3,923
118	🇸🇰 Slovakia	13	27	8.0	7.4	5,756	25,217	292	193	883	2,516
119	🇬🇮 Gibraltar	1	21	7.9	7.9	7,673	16,102	386	3,982	632	2,282
120	🇲🇹 Malta	31	56	7.6	7.6	3,580	28,190	50	534	761	3,930
121	🇨🇮 Côte d'Ivoire	2	19	7.6	7.6	14,132	33,563	176	1,491	927	4,248
122	🇱🇺 Luxembourg	12	25	7.0	7.0	8,621	16,715	475	316	917	1,657
123	🇰🇼 Kuwait	2	21	6.9	6.9	2,718	27,860	27	74	1,073	2,699
124	🇱🇷 Liberia	11	43	6.8	6.8	28,778	18,445	121	10,027	983	5,412
125	🇸🇷 Suriname	3	19	6.7	6.7	3,821	22,472	175	713	757	3,641
126	🇿🇼 Zimbabwe	4	26	6.3	6.3	18,762	28,919	112	147	1,302	6,236
127	🇦🇬 Antigua and Barbuda	9	21	6.1	6.1	4,896	12,483	52	26	279	2,036
128	🇵🇾 Paraguay	8	18	5.9	5.9	6,265	21,646	300	3,071	471	4,110
129	🇪🇷 Eritrea	11	29	5.8	4.6	15,426	7,443	3,418	138	286	1,732
130	🇻🇨 Saint Vincent and the Grenadines	11	25	5.7	5.7	4,066	8,161	23	30	549	1,879
131	🇧🇯 Benin	10	16	5.6	5.6	21,948	11,234	325	13,514	293	2,270
132	🇬🇺 Guam	10	44	5.5	4.5	511	17,057	14	33	154	3,148
133	🇦🇫 Afghanistan	8	26	5.3	5.3	13,045	18,705	1,027	728	890	3,491
134	🇨🇫 Central African Republic	3	18	5.3	5.3	38,834	7,971	375	15,448	187	2,454
135	🇱🇮 Liechtenstein	16	59	5.2	5.2	1,175	9,182	176	210	434	1,491
136	🇹🇳 Tunisia	5	15	5.1	5.1	10,132	26,250	173	676	507	3,371
137	🇸🇳 Senegal	6	21	5.1	4.3	12,027	13,457	159	3,041	466	2,450
138	🇹🇹 Trinidad and Tobago	3	5	4.6	4.6	1,723	9,873	3	11	229	1,814
139	🇦🇩 Andorra	8	45	4.6	4.6	4,311	14,526	84	1,199	649	2,824
140	🇳🇮 Nicaragua	8	17	4.6	4.6	6,405	10,464	600	2,780	483	3,249
141	🇬🇬 Guernsey	8	24	4.5	3.1	1,575	6,630	113	166	187	986
142	🇲🇴 Macau	2	15	4.5	4.5	3,426	6,072	47	658	917	986
143	🇰🇵 North Korea	1	19	4.4	4.4	15,858	4,007	1,874	103	934	773
144	🇨🇷 Costa Rica	10	19	4.3	4.3	4,443	10,745	253	769	450	2,188
145	🇲🇬 Madagascar	3	11	4.3	4.3	23,805	7,053	2,068	2,796	1,017	3,460
146	🇵🇫 French Polynesia	11	17	4.3	4.3	2,054	12,019	210	985	196	2,632
147	🇰🇾 Cayman Islands	5	17	4.2	4.2	764	10,965	83	36	319	2,040
148	🇬🇵 Guadeloupe	11	30	4.1	4.1	1,466	15,765	41	118	285	1,953
149	🇦🇴 Angola	2	17	4.1	4.1	11,795	15,013	70	959	1,047	3,617
150	🇿🇲 Zambia	4	13	3.7	3.7	7,413	7,263	410	96	532	2,294
151	🇷🇼 Rwanda	7	14	3.5	2.9	13,643	5,940	643	4,842	327	1,609
152	🇬🇱 Greenland	3	11	3.5	3.5	1,601	4,186	42	4	186	967
153	🇸🇲 San Marino	4	13	3.5	3.5	1,033	9,251	11	245	155	894
154	🇨🇩 Congo (Democratic Republic)	3	16	3.3	3.2	10,412	8,123	66	2,680	289	1,549
155	🇨🇲 Cameroon	9	16	3.2	3.2	10,497	12,297	42	3,015	200	1,731
156	🇫🇯 Fiji	3	9	3.2	3.2	2,210	7,945	6	24	454	1,671
157	🇱🇨 Saint Lucia	7	20	3.1	3.1	2,637	7,941	17	45	591	2,042
158	🇬🇦 Gabon	6	17	3.0	3.0	4,951	7,125	33	81	182	1,966
159	🇭🇹 Haiti	1	6	2.9	2.9	8,800	4,956	45	1,599	432	2,797
160	🇩🇴 Dominican Republic	7	9	2.9	2.9	3,892	7,216	172	1,900	298	1,365
161	🇸🇴 Somalia	4	15	2.8	2.8	8,102	4,946	21	172	881	2,647
162	🇲🇺 Mauritius	7	13	2.8	2.8	3,727	7,593	76	1,021	457	1,523
163	🇧🇲 Bermuda	4	5	2.7	2.7	665	3,426	24	747	189	715
164	🇳🇨 New Caledonia	1	5	2.6	2.6	678	6,105	49	42	103	1,410
165	🇾🇪 Yemen	3	12	2.5	2.5	7,733	10,455	96	2,542	348	2,854
166	🇰🇲 Comoros	3	5	2.5	2.5	3,655	5,007	55	1,568	87	776
167	🇧🇼 Botswana	4	9	2.3	2.3	1,646	5,688	5	10	174	1,109
168	🇯🇪 Jersey	10	18	2.2	2.2	908	4,816	101	68	148	969
169	🇻🇬 British Virgin Islands	7	17	2.2	2.2	693	5,394	3	49	101	1,205
170	🇰🇳 Saint Kitts and Nevis	3	9	2.1	2.1	1,807	3,326	26	25	174	746
171	🇼🇸 Samoa	1	7	2.0	2.0	600	4,406	0	0	265	720
172	🇷🇪 Réunion	4	6	2.0	2.0	1,254	5,554	31	90	107	643
173	🇹🇯 Tajikistan	2	9	1.9	1.9	2,684	11,087	30	72	355	878
174	🇸🇸 South Sudan	1	5	1.9	1.9	8,871	4,502	76	3,089	168	1,511
175	🇳🇦 Namibia	4	10	1.8	1.8	1,760	5,176	18	11	95	936
176	🇫🇴 Faroe Islands	21	36	1.8	1.8	294	2,928	12	4	125	472
177	🇬🇫 French Guiana	1	3	1.7	1.7	850	7,399	59	389	71	1,349
178	🇮🇲 Isle of Man	8	19	1.6	1.6	1,428	7,437	24	689	210	1,162
179	🇧🇹 Bhutan	2	6	1.6	1.6	2,519	4,667	20	79	120	953
180	🇩🇯 Djibouti	3	5	1.5	1.5	764	1,675	5	116	245	1,383
181	🇱🇸 Lesotho	10	15	1.5	1.5	1,858	2,512	5	11	95	648
182	🇬🇩 Grenada	4	8	1.4	1.4	585	1,858	8	43	97	617
183	🇲🇫 Saint Martin	6	9	1.3	1.3	468	4,583	13	132	100	1,182
184	🇧🇮 Burundi	6	9	1.3	1.3	5,894	2,341	436	314	173	493
185	🇧🇧 Barbados	3	4	1.3	1.3	1,700	3,063	42	7	179	514
186	🇲🇿 Mozambique	5	8	1.3	1.3	2,000	1,892	34	150	119	465
187	🇲🇷 Mauritania	1	3	1.3	1.3	2,055	4,774	74	116	38	980
188	🇪🇭 Western Sahara	3	4	1.3	1.3	767	2,863	34	101	56	578
189	🇹🇬 Togo	6	11	1.2	1.2	3,040	1,492	59	1,735	55	372
190	🇵🇬 Papua New Guinea	2	6	1.2	1.2	1,133	1,765	0	0	217	444
191	🇲🇸 Montserrat	1	3	1.2	1.2	6,844	1,223	274	6,156	393	892
192	🇦🇸 American Samoa	9	13	1.1	1.1	269	2,750	5	1	127	1,200
193	🇧🇸 Bahamas	2	5	1.1	1.1	358	1,866	21	24	63	294
194	🇹🇨 Turks and Caicos Islands	2	4	1.0	1.0	102	1,577	0	2	23	218
195	🇻🇺 Vanuatu	1	5	1.0	1.0	1,064	1,485	14	3	157	613
196	🇹🇩 Chad	1	4	1.0	1.0	4,357	3,011	75	2,048	52	764
197	🇻🇮 United States Virgin Islands	3	5	1.0	1.0	399	2,875	0	7	72	787
198	🇹🇴 Tonga	1	2	0.9	0.4	95	1,134	1	1	41	279
199	🇲🇱 Mali	1	2	0.9	0.9	5,096	2,540	29	3,350	174	705
200	🇳🇷 Nauru	11	18	0.9	0.9	329	1,072	4	343	49	571
201	🇧🇫 Burkina Faso	1	5	0.8	0.8	3,127	3,133	325	2,983	55	463
202	🇸🇽 Sint Maarten	4	5	0.8	0.8	632	3,217	7	42	67	621
203	🇦🇼 Aruba	2	3	0.8	0.8	312	2,372	6	9	44	272
204	🇨🇼 Curaçao	1	3	0.8	0.8	32	1,756	0	0	15	197
205	🇦🇽 Åland Islands	1	1	0.8	0.8	112	770	26	4	11	100
206	🇲🇭 Marshall Islands	3	3	0.8	0.8	114	340	0	0	13	95
207	🇧🇳 Brunei	3	4	0.8	0.8	74	1,456	1	4	27	146
208	🇸🇯 Svalbard and Jan Mayen	1	3	0.7	0.7	60	529	5	5	39	173
209	🇧🇿 Belize	3	5	0.7	0.7	409	1,622	61	91	49	526
210	🇧🇶 Bonaire, Sint Eustatius and Saba	1	1	0.7	0.7	237	1,967	21	82	27	734
211	🇬🇶 Equatorial Guinea	1	5	0.7	0.7	828	2,165	12	9	51	379
212	🇫🇲 Micronesia	3	3	0.6	0.6	165	680	0	0	6	148
213	🇳🇪 Niger	1	2	0.6	0.6	2,932	1,126	66	270	7	128
214	🇵🇼 Palau	4	5	0.6	0.6	23	548	3	3	9	127
215	🇦🇮 Anguilla	4	5	0.6	0.6	57	788	0	1	15	178
216	🇸🇩 Sudan	2	6	0.6	0.6	1,101	3,448	8	181	141	663
217	🇧🇱 Saint Barthélemy	2	3	0.5	0.5	218	1,485	0	225	22	525
218	🇲🇵 Northern Mariana Islands	2	5	0.5	0.5	109	1,819	3	0	20	304
219	🇲🇻 Maldives	2	4	0.5	0.5	1,708	768	57	1,557	36	187
220	🇵🇲 Saint Pierre and Miquelon	1	1	0.4	0.4	76	1,203	2	5	18	382
221	🇸🇿 Eswatini	3	3	0.4	0.4	653	890	1	1	37	162
222	🇹🇻 Tuvalu	1	2	0.4	0.4	118	68	3	211	6	40
223	🇫🇰 Falkland Islands	1	4	0.3	0.3	37	82	0	0	18	93
224	🇨🇰 Cook Islands	2	5	0.3	0.3	88	339	7	125	10	74
225	🇸🇨 Seychelles	3	3	0.3	0.3	280	620	10	12	42	176
226	🇸🇭 Saint Helena	2	3	0.3	0.3	60	347	0	6	28	107
227	🇩🇲 Dominica	1	2	0.3	0.3	192	466	1	8	19	144
228	🇹🇱 Timor-Leste	5	5	0.3	0.3	278	125	0	84	31	59
229	🇳🇫 Norfolk Island	1	2	0.3	0.3	4	16	0	0	3	8
230	🇨🇻 Cabo Verde	2	2	0.2	0.2	301	567	17	16	30	145
231	🇲🇶 Martinique	3	4	0.2	0.2	99	553	1	6	7	69
232	🇸🇹 Sao Tome and Principe	1	1	0.2	0.2	525	205	20	204	6	91
233	🇬🇼 Guinea-Bissau	1	1	0.2	0.2	776	487	4	40	23	94
234	🇾🇹 Mayotte	2	2	0.2	0.2	125	264	0	25	20	50
235	🇸🇧 Solomon Islands	1	2	0.1	0.1	168	390	0	0	38	135
236	🇨🇽 Christmas Island	1	1	0.1	0.1	57	132	0	1	7	32
237	🇳🇺 Niue	3	3	0.1	0.1	0	39	0	0	3	9
238	🇰🇮 Kiribati	2	2	0.1	0.1	74	102	5	17	8	31

Top 200 cities by footage

#	City	Country	Videos	Footage (h)	Analysed (h)	Persons	Cars	Bicycles	Motorcycles	Buses	Trucks
1	New York, NY	🇺🇸 United States	4,179	2,280.4	1,470.3	1,832,920	5,726,186	132,939	66,353	186,612	815,162
2	Chicago, IL	🇺🇸 United States	1,896	1,713.9	1,701.9	713,628	6,574,579	39,424	12,501	89,547	794,980
3	London	🇬🇧 United Kingdom	2,038	1,363.4	1,180.9	2,632,307	2,672,124	249,529	145,538	236,802	467,612
4	Los Angeles, CA	🇺🇸 United States	2,047	1,164.8	1,089.2	647,916	4,613,976	33,409	17,639	66,525	600,247
5	Hanoi	🇻🇳 Vietnam	2,151	876.0	837.5	2,424,080	1,676,026	63,948	2,301,030	100,554	217,041
6	Manila, Metro Manila	🇵🇭 Philippines	2,179	793.6	793.6	1,234,980	1,481,787	36,913	636,588	168,528	358,530
7	Sydney, NSW	🇦🇺 Australia	1,432	682.0	682.0	399,269	2,822,235	29,697	34,706	68,127	448,501
8	Saint Petersburg, Saint Petersburg	🇷🇺 Russia	689	566.0	566.0	297,884	1,534,337	6,959	1,923	68,492	168,517
9	Seoul	🇰🇷 South Korea	294	543.0	542.9	261,221	1,138,257	14,008	41,210	84,270	97,020
10	Jakarta	🇮🇩 Indonesia	1,674	475.0	472.5	838,608	1,089,241	6,600	740,038	63,212	246,701
11	Paris	🇫🇷 France	981	437.9	410.3	597,476	1,126,516	68,949	88,940	31,323	123,781
12	Toronto, ON	🇨🇦 Canada	885	375.3	373.6	491,485	1,301,773	44,351	12,820	25,461	138,084
13	Yangon	🇲🇲 Myanmar	488	357.1	357.1	365,295	1,623,845	41,365	20,924	98,139	252,165
14	Melbourne, VIC	🇦🇺 Australia	671	329.8	312.1	213,303	1,218,183	13,359	11,063	21,201	131,170
15	Moscow, Moscow	🇷🇺 Russia	505	327.5	327.5	214,143	1,172,898	4,806	3,583	36,218	152,129
16	Dubai	🇦🇪 United Arab Emirates	779	320.6	260.2	84,298	693,915	2,877	14,020	37,278	91,282
17	Philadelphia, PA	🇺🇸 United States	564	316.6	297.7	146,579	1,071,550	8,386	4,177	14,456	146,815
18	Bangkok	🇹🇭 Thailand	625	313.6	293.3	397,188	874,639	6,930	276,661	40,549	168,457
19	Cotabato, Maguindanao del Norte	🇵🇭 Philippines	739	301.4	287.0	762,978	595,688	15,892	535,140	97,078	412,619
20	Sao Paulo, SP	🇧🇷 Brazil	566	251.1	249.6	248,005	748,836	11,517	63,161	21,444	89,873
21	København	🇩🇰 Denmark	534	246.1	239.1	384,482	757,903	198,371	8,060	19,704	59,110
22	Kyiv, Kyiv	🇺🇦 Ukraine	576	242.4	234.5	167,637	993,383	4,849	2,489	39,574	125,363
23	Kuala Lumpur, KUL	🇲🇾 Malaysia	346	237.3	159.1	97,861	488,945	727	50,334	10,895	39,393
24	Warsaw	🇵🇱 Poland	391	236.9	218.0	157,200	715,347	14,247	5,484	23,846	54,856
25	Berlin	🇩🇪 Germany	460	219.4	210.4	264,273	873,904	83,063	20,563	27,706	103,961
26	Buenos Aires, CABA	🇦🇷 Argentina	286	216.7	216.1	326,712	799,840	31,682	35,728	28,015	80,544
27	Townsville, QLD	🇦🇺 Australia	436	207.5	6.2	967	16,946	61	115	259	3,737
28	Detroit, MI	🇺🇸 United States	559	206.6	199.2	26,440	608,888	1,203	814	5,832	94,345
29	Guangzhou, Guangdong	🇨🇳 China	245	205.8	205.8	133,101	614,446	13,130	49,544	36,604	76,299
30	Rio de Janeiro, RJ	🇧🇷 Brazil	465	203.6	146.7	163,951	489,625	11,304	28,259	23,330	48,225
31	Montreal, QC	🇨🇦 Canada	472	203.2	203.2	122,326	641,567	13,433	3,302	11,036	50,954
32	Phnom Penh	🇰🇭 Cambodia	387	201.9	201.3	571,210	522,075	12,618	588,950	35,841	168,858
33	Helsinki	🇫🇮 Finland	284	198.6	196.7	112,793	429,462	10,680	1,721	19,525	49,127
34	Prague	🇨🇿 Czechia	269	196.7	187.1	209,912	508,633	5,088	4,832	20,302	45,692
35	Miami, FL	🇺🇸 United States	467	186.8	144.8	99,616	477,642	5,820	7,990	6,488	60,965
36	Budapest	🇭🇺 Hungary	520	183.1	164.4	307,084	1,099,775	14,395	8,242	50,555	124,155
37	Vienna	🇦🇹 Austria	766	181.8	173.5	109,291	489,394	12,122	12,691	20,406	76,036
38	Kampala	🇺🇬 Uganda	471	178.2	68.2	404,441	163,845	10,066	215,904	15,266	53,395
39	Stockholm	🇸🇪 Sweden	582	178.0	169.6	95,871	304,628	13,295	1,858	12,926	34,293
40	Tokyo, Tokyo	🇯🇵 Japan	259	172.5	170.2	259,142	336,484	23,710	11,812	32,564	92,696
41	Istanbul	🇹🇷 Türkiye	317	172.5	128.9	170,292	547,699	1,409	20,210	34,776	77,396
42	Eindhoven	🇳🇱 Netherlands	329	171.2	171.2	162,292	435,704	111,943	15,541	10,050	39,732
43	Yogyakarta	🇮🇩 Indonesia	298	171.0	168.9	533,348	345,857	7,646	549,320	13,370	70,843
44	Perth, WA	🇦🇺 Australia	418	170.9	99.6	18,872	272,695	1,209	1,157	5,095	45,650
45	Vancouver, BC	🇨🇦 Canada	711	170.7	40.3	38,811	139,606	4,311	746	2,763	18,614
46	Bucharest	🇷🇴 Romania	331	169.6	165.0	130,061	857,986	4,801	4,188	19,749	68,862
47	Nashville, TN	🇺🇸 United States	856	168.7	51.5	37,893	145,831	497	355	2,234	22,416
48	Singapore	🇸🇬 Singapore	417	163.8	152.6	84,721	384,215	5,235	20,853	33,904	61,744
49	Atlanta, GA	🇺🇸 United States	373	159.7	154.7	32,717	423,664	1,270	943	4,958	81,366
50	Sacramento, CA	🇺🇸 United States	413	159.5	158.0	17,540	614,371	1,273	972	6,089	99,491
51	Houston, TX	🇺🇸 United States	238	158.9	37.4	9,717	100,673	315	119	1,383	16,586
52	Winnipeg, MB	🇨🇦 Canada	570	154.1	59.1	21,126	204,135	1,231	425	3,261	29,636
53	Oldenburg	🇩🇪 Germany	173	153.0	151.4	82,063	236,585	43,991	3,357	7,435	22,631
54	Tyumen, Tyumen oblast	🇷🇺 Russia	306	147.5	147.5	95,141	482,654	2,000	538	19,971	49,974
55	Sofia	🇧🇬 Bulgaria	464	140.6	132.4	69,793	471,103	1,321	911	15,195	37,361
56	Washington, DC	🇺🇸 United States	251	139.1	121.3	86,395	421,265	6,710	2,408	8,490	49,484
57	San Jose, CA	🇺🇸 United States	288	134.9	134.9	25,981	655,825	1,736	1,111	6,901	84,585
58	Tehran	🇮🇷 Iran	313	132.3	69.6	52,110	380,653	1,291	19,820	5,565	30,477
59	Den Haag	🇳🇱 Netherlands	270	131.7	27.8	19,638	98,650	13,069	3,301	2,360	9,188
60	Dublin	🇮🇪 Ireland	426	129.9	114.9	100,657	288,687	10,200	2,042	12,828	35,654
61	San Francisco, CA	🇺🇸 United States	194	128.0	128.0	173,821	462,468	8,962	6,884	10,965	54,916
62	Maracaibo, ZU	🇻🇪 Venezuela	535	125.7	119.2	82,775	333,647	4,471	30,736	6,100	64,464
63	Anchorage, AK	🇺🇸 United States	347	125.1	122.7	5,861	199,273	276	177	2,781	45,133
64	Dallas, TX	🇺🇸 United States	295	124.9	120.6	13,963	239,842	365	345	2,841	38,376
65	Busan	🇰🇷 South Korea	143	121.7	121.7	69,400	348,689	2,527	8,548	12,381	39,265
66	Las Vegas, NV	🇺🇸 United States	172	120.0	111.0	60,345	321,846	573	926	5,824	41,881
67	Oslo	🇳🇴 Norway	321	119.9	115.0	74,207	274,151	6,511	1,643	11,909	33,429
68	Volsk, Saratov oblast	🇷🇺 Russia	525	116.5	89.1	40,112	155,727	299	241	4,104	17,041
69	Baltimore, MD	🇺🇸 United States	325	116.5	79.5	22,656	273,572	484	426	3,958	37,368
70	Portland, OR	🇺🇸 United States	185	114.7	90.8	57,267	236,884	3,756	873	4,378	28,336
71	Amsterdam	🇳🇱 Netherlands	207	110.9	108.3	256,026	227,672	182,306	28,327	11,015	31,177
72	Phoenix, AZ	🇺🇸 United States	228	106.8	44.7	9,484	121,463	451	293	1,494	24,917
73	Hong Kong	🇭🇰 Hong Kong	200	99.6	56.2	103,516	101,745	1,186	2,868	18,631	18,363
74	Mumbai, MH	🇮🇳 India	263	98.5	98.0	258,652	359,189	6,469	102,036	19,407	76,972
75	Seattle, WA	🇺🇸 United States	206	97.5	2.5	3,081	6,736	198	53	164	546
76	Darwin, NT	🇦🇺 Australia	310	93.9	44.6	12,771	105,322	454	933	1,835	20,261
77	Cairo	🇪🇬 Egypt	206	91.6	39.0	62,617	225,545	503	6,961	9,939	25,647
78	Cape Town	🇿🇦 South Africa	237	89.8	82.8	57,449	258,070	1,333	3,830	7,821	39,978
79	Kermanshah	🇮🇷 Iran	247	89.6	88.5	59,387	577,645	573	3,505	3,746	42,779
80	Minsk	🇧🇾 Belarus	115	84.7	26.9	19,992	117,544	481	235	5,100	13,656
81	München	🇩🇪 Germany	170	82.1	81.8	84,961	289,323	27,266	4,465	10,061	33,590
82	Serrekunda	🇬🇲 Gambia	273	78.5	77.0	195,185	282,496	10,698	17,296	6,850	54,231
83	Amman	🇯🇴 Jordan	262	78.5	61.9	37,103	303,513	107	881	5,905	32,768
84	Ho Chi Minh City	🇻🇳 Vietnam	141	74.9	60.0	206,416	87,338	3,361	211,366	7,776	20,792
85	Sylhet	🇧🇩 Bangladesh	194	74.5	74.5	329,931	80,917	59,886	57,802	27,021	65,202
86	Damascus	🇸🇾 Syria	163	72.4	65.6	104,871	306,275	4,564	11,498	10,456	44,470
87	Brisbane, QLD	🇦🇺 Australia	188	71.7	67.5	31,938	197,023	1,317	1,928	4,513	35,310
88	Denver, CO	🇺🇸 United States	204	71.6	58.4	17,410	134,277	1,325	581	2,038	21,062
89	Gamagori, Aichi	🇯🇵 Japan	271	71.6	68.8	13,153	161,444	3,182	1,487	6,737	39,329
90	Ulaanbaatar	🇲🇳 Mongolia	127	70.2	70.2	91,683	343,965	1,881	1,671	9,239	26,270
91	San Diego, CA	🇺🇸 United States	94	68.2	68.2	55,195	256,391	2,337	1,874	3,415	34,968
92	Karachi, SD	🇵🇰 Pakistan	147	67.1	63.2	209,864	263,480	6,697	105,138	17,546	60,258
93	Semarang	🇮🇩 Indonesia	322	61.7	61.7	92,280	98,757	1,624	92,742	6,749	32,420
94	Auckland	🇳🇿 New Zealand	142	61.5	61.5	47,632	190,328	2,563	1,657	6,810	21,915
95	Tallinn	🇪🇪 Estonia	288	61.2	60.4	32,061	174,551	1,075	374	8,149	15,986
96	Denpasar	🇮🇩 Indonesia	28	60.6	60.6	239,875	103,387	4,156	284,862	3,168	25,458
97	Calgary, AB	🇨🇦 Canada	170	60.0	60.0	22,254	170,303	954	248	2,358	25,440
98	Ninh Binh	🇻🇳 Vietnam	197	58.2	55.8	65,871	59,322	6,843	76,410	3,702	15,669
99	Delhi, DL	🇮🇳 India	165	57.8	57.6	98,456	167,913	3,782	39,965	18,744	31,632
100	Beijing, Beijing	🇨🇳 China	130	57.2	57.2	47,041	175,888	11,385	11,687	10,252	8,869
101	Aleppo	🇸🇾 Syria	155	57.0	50.6	117,001	233,299	1,749	4,638	6,259	37,306
102	Catania	🇮🇹 Italy	144	56.5	56.5	17,419	140,478	199	8,775	1,665	6,724
103	Shanghai, Shanghai	🇨🇳 China	79	56.2	40.0	56,793	93,459	6,533	15,032	6,043	6,679
104	Boston, MA	🇺🇸 United States	110	54.6	35.0	58,107	106,705	2,760	770	2,398	14,554
105	Milan	🇮🇹 Italy	104	54.2	52.9	78,811	204,153	10,266	16,666	3,909	16,019
106	Orlando, FL	🇺🇸 United States	113	52.1	52.1	17,140	156,895	1,255	652	2,565	20,453
107	Shenzhen, Guangdong	🇨🇳 China	143	51.6	46.2	39,411	125,997	4,331	13,374	9,926	14,122
108	Honolulu, HI	🇺🇸 United States	118	49.7	24.0	15,995	65,301	960	669	1,447	8,588
109	Ben Tre	🇻🇳 Vietnam	99	49.0	44.6	151,402	32,419	4,750	148,535	3,321	12,171
110	Santiago	🇨🇱 Chile	86	48.3	2.6	3,883	9,327	174	360	647	1,271
111	Barcelona	🇪🇸 Spain	100	48.3	46.8	78,935	107,455	7,795	31,623	6,632	14,301
112	Mexico City, CDMX	🇲🇽 Mexico	75	47.2	42.4	55,326	133,085	2,590	4,691	4,081	12,249
113	Kolkata, WB	🇮🇳 India	150	46.6	45.9	131,183	78,970	21,787	35,016	19,196	35,944
114	Toyohashi, Aichi	🇯🇵 Japan	170	45.7	42.6	11,140	98,404	2,794	876	4,731	25,546
115	Lviv, Lviv oblast	🇺🇦 Ukraine	150	45.0	44.2	62,463	131,483	966	381	8,372	17,761
116	Mashhad	🇮🇷 Iran	251	44.6	41.4	16,932	149,750	402	3,882	2,166	6,353
117	Ninohe, Iwate	🇯🇵 Japan	247	44.4	44.1	7,161	81,846	390	333	5,640	25,418
118	Sarajevo	🇧🇦 Bosnia and Herzegovina	159	44.1	44.0	55,695	168,303	579	442	4,571	11,698
119	Baku	🇦🇿 Azerbaijan	100	43.3	38.3	38,449	217,459	527	1,060	5,529	18,189
120	Colombo	🇱🇰 Sri Lanka	107	43.1	42.8	66,996	94,312	843	21,806	15,645	35,260
121	Adelaide, SA	🇦🇺 Australia	123	42.5	42.5	9,589	144,333	917	567	2,813	17,586
122	Addis Ababa	🇪🇹 Ethiopia	68	42.3	40.4	115,291	123,221	474	1,806	9,115	33,185
123	Surabaya	🇮🇩 Indonesia	106	41.1	37.9	76,546	62,767	1,232	66,304	1,974	10,890
124	Nam Dinh	🇻🇳 Vietnam	128	40.9	35.6	67,264	46,449	5,821	79,561	1,689	8,133
125	Zaozhuang, Shandong	🇨🇳 China	123	40.0	39.9	24,179	141,876	1,348	14,078	6,007	20,080
126	Doha	🇶🇦 Qatar	78	39.5	39.5	10,295	115,819	407	1,233	4,840	15,020
127	Zürich, ZH	🇨🇭 Switzerland	76	39.4	39.4	57,294	87,354	10,073	5,345	4,264	11,207
128	Austin, TX	🇺🇸 United States	69	39.0	33.8	19,552	105,605	1,457	790	1,453	19,517
129	Rome	🇮🇹 Italy	76	38.7	36.7	63,696	163,290	2,244	12,898	3,499	10,965
130	Jersey City, NJ	🇺🇸 United States	104	38.6	35.1	15,146	123,022	1,190	691	2,027	13,948
131	Madrid	🇪🇸 Spain	76	38.2	37.4	55,365	102,996	1,483	8,631	5,232	9,974
132	Asaba, Delta	🇳🇬 Nigeria	84	38.0	9.7	19,671	27,231	41	622	4,592	10,524
133	Wellington	🇳🇿 New Zealand	125	37.3	34.8	18,816	110,183	987	1,074	3,130	12,874
134	Nicosia	🇨🇾 Cyprus	127	37.3	37.0	10,889	137,042	466	2,896	2,334	16,776
135	Lisbon	🇵🇹 Portugal	70	37.3	37.3	31,376	124,157	1,072	3,710	3,362	11,946
136	Kharkiv, Kharkiv oblast	🇺🇦 Ukraine	80	37.2	35.1	24,393	105,583	322	259	3,611	12,058
137	Brussels	🇧🇪 Belgium	82	37.1	37.1	47,072	118,563	6,812	2,465	5,031	14,930
138	Freetown	🇸🇱 Sierra Leone	91	36.9	36.7	102,242	41,711	676	21,742	3,869	18,994
139	Johor Bahru, JHR	🇲🇾 Malaysia	39	36.8	29.1	10,897	123,426	179	7,710	2,194	9,837
140	Kathmandu	🇳🇵 Nepal	112	35.7	35.7	220,231	70,843	5,498	128,573	10,501	19,109
141	Zagreb	🇭🇷 Croatia	94	35.7	35.7	27,817	147,683	2,835	958	3,150	9,856
142	Lagos, Lagos	🇳🇬 Nigeria	73	35.1	34.3	168,544	111,389	1,374	17,402	9,296	28,192
143	Genoa	🇮🇹 Italy	73	35.1	34.3	47,640	135,749	928	39,880	3,761	13,035
144	Chennai, TN	🇮🇳 India	47	34.8	34.0	90,190	44,683	1,507	39,928	14,681	24,217
145	Riga	🇱🇻 Latvia	98	34.4	33.9	24,183	98,511	1,558	704	4,078	11,828
146	Bengaluru, KA	🇮🇳 India	97	34.3	32.5	80,872	75,362	1,237	51,157	12,119	25,982
147	Belgrade	🇷🇸 Serbia	106	34.2	32.1	31,869	142,501	446	950	5,003	10,565
148	Thessaloniki	🇬🇷 Greece	81	33.8	33.4	43,803	174,434	1,217	12,889	3,364	13,658
149	Tampa, FL	🇺🇸 United States	52	32.4	32.4	10,866	91,150	633	396	983	12,910
150	Birmingham	🇬🇧 United Kingdom	119	32.3	31.4	16,010	111,770	452	331	3,145	12,970
151	Riyadh	🇸🇦 Saudi Arabia	79	32.3	18.7	2,938	111,390	52	182	2,081	12,310
152	Rotterdam	🇳🇱 Netherlands	80	31.0	30.7	29,958	89,885	15,849	3,300	2,304	8,778
153	Goiânia, GO	🇧🇷 Brazil	69	30.8	28.8	15,865	85,432	807	8,619	1,146	8,190
154	Edmonton, AB	🇨🇦 Canada	56	30.4	30.4	7,755	105,060	402	217	1,530	14,082
155	Chengdu, Sichuan	🇨🇳 China	57	30.3	27.9	17,404	115,943	2,479	5,513	4,291	8,216
156	Tbilisi	🇬🇪 Georgia	75	29.6	29.6	23,420	148,631	124	1,190	4,914	15,300
157	Kansas City, KS	🇺🇸 United States	53	29.1	27.8	3,482	71,522	94	70	857	11,488
158	Toshkent	🇺🇿 Uzbekistan	86	28.7	28.2	20,874	208,217	200	84	3,287	13,238
159	Milton Keynes	🇬🇧 United Kingdom	50	27.7	27.7	1,585	65,583	173	187	752	5,973
160	Pittsburgh, PA	🇺🇸 United States	58	27.6	23.9	13,679	52,589	580	244	1,098	7,870
161	Annaba	🇩🇿 Algeria	40	27.1	7.9	17,553	39,520	114	866	720	3,975
162	Yerevan	🇦🇲 Armenia	89	27.1	20.1	18,207	140,848	119	315	4,164	13,759
163	Lahore, PB	🇵🇰 Pakistan	89	26.9	11.7	40,182	48,905	1,440	24,541	1,482	8,359
164	Omaha, NE	🇺🇸 United States	82	26.6	26.1	3,443	86,561	181	211	695	12,840
165	Manchester	🇬🇧 United Kingdom	61	26.4	26.2	20,158	75,360	981	268	2,401	8,283
166	Havana	🇨🇺 Cuba	129	26.2	26.2	98,261	46,170	2,797	13,826	3,375	12,455
167	Cherepovets, Vologda oblast	🇷🇺 Russia	176	25.0	24.5	26,836	77,993	65	25	4,097	10,486
168	Hangzhou, Zhejiang	🇨🇳 China	35	24.9	8.8	6,905	25,486	414	2,780	1,186	1,999
169	Lima, LI	🇵🇪 Peru	37	24.7	24.3	44,246	110,551	4,244	6,512	4,856	12,370
170	Thiruvananthapuram, KL	🇮🇳 India	74	24.7	24.5	64,978	58,448	1,023	61,737	6,718	22,364
171	Daejeon	🇰🇷 South Korea	61	24.3	24.0	9,720	79,368	1,046	1,141	2,795	5,982
172	Frankfurt	🇩🇪 Germany	20	24.3	0.0	–	–	–	–	–	–
173	Vilnius	🇱🇹 Lithuania	60	24.2	24.2	27,749	74,129	2,322	625	2,345	4,959
174	Samarkand	🇺🇿 Uzbekistan	81	23.9	23.5	24,145	140,637	407	234	2,900	13,034
175	Islamabad, ICT	🇵🇰 Pakistan	57	23.9	23.6	30,443	75,869	1,256	14,131	2,143	9,602
176	Stavropol, Stavropol krai	🇷🇺 Russia	79	23.9	23.9	13,759	93,659	262	83	3,365	8,364
177	New Orleans, LA	🇺🇸 United States	34	23.7	21.8	25,997	60,205	1,223	356	1,254	11,443
178	Brampton, ON	🇨🇦 Canada	75	23.1	22.9	5,341	114,696	162	124	1,566	12,067
179	Charleston, SC	🇺🇸 United States	38	23.0	23.0	25,616	60,499	934	308	853	13,020
180	Sendai, Miyagi	🇯🇵 Japan	58	22.7	22.5	6,350	64,201	898	511	3,185	13,403
181	Ōsaka, Ōsaka	🇯🇵 Japan	52	22.5	21.2	31,030	45,743	7,723	1,642	3,074	9,491
182	Porto	🇵🇹 Portugal	27	22.4	22.4	31,466	65,534	360	1,680	1,831	6,603
183	San Antonio, TX	🇺🇸 United States	26	22.2	11.1	9,257	23,720	350	146	424	4,357
184	Ottawa, ON	🇨🇦 Canada	38	22.1	22.1	29,945	61,718	3,099	432	1,485	6,471
185	Kyoto, Kyōto	🇯🇵 Japan	45	21.9	19.7	28,572	42,313	3,422	2,068	4,455	10,468
186	Taipei	🇹🇼 Taiwan	50	21.8	19.6	27,758	55,496	1,764	22,112	3,229	6,334
187	Sapporo, Hokkaidō	🇯🇵 Japan	62	21.8	20.8	7,971	39,771	526	92	2,070	4,873
188	Cheongju	🇰🇷 South Korea	65	21.7	21.4	6,303	66,092	316	676	2,069	7,761
189	Skopje	🇲🇰 North Macedonia	42	21.1	18.9	21,867	79,024	1,435	757	2,149	6,441
190	Chiang Mai	🇹🇭 Thailand	53	21.0	21.0	16,101	44,903	508	16,273	1,513	13,247
191	Halifax, NS	🇨🇦 Canada	82	20.8	20.8	9,402	54,252	342	149	1,004	5,668
192	Chișinău	🇲🇩 Moldova	73	20.8	19.7	23,126	103,674	170	173	4,552	10,888
193	Medellín, ANT	🇨🇴 Colombia	53	20.8	10.0	24,671	29,479	718	11,404	1,774	3,569
194	Beirut	🇱🇧 Lebanon	68	20.5	20.5	17,401	112,506	108	6,327	1,864	10,267
195	Monaco	🇲🇨 Monaco	45	20.3	20.3	38,846	32,963	1,033	13,199	1,598	4,440
196	Pattaya	🇹🇭 Thailand	36	19.9	19.0	30,427	49,365	286	22,521	3,284	13,820
197	Manaus, AM	🇧🇷 Brazil	75	19.8	19.8	20,325	76,162	214	10,206	2,301	10,836
198	Ankara	🇹🇷 Türkiye	47	19.3	19.3	16,759	98,754	48	858	4,915	9,692
199	Malibu, CA	🇺🇸 United States	34	19.2	16.7	4,135	78,658	276	452	786	14,809
200	Nice	🇫🇷 France	32	19.0	18.2	33,313	50,774	3,540	7,408	1,300	5,547

Citation and usage of code

If you use this work for academic work please cite the following paper:

Alam, M. S., Bazilinska, O., & Bazilinskyy, P. (2026). A global dataset of continuous urban dashcam driving. arXiv preprint arXiv:2604.01044. Under review. Available at https://arxiv.org/abs/2604.01044

What the dataset contains, how it was collected and its known biases are described in the datasheet. Citation metadata is in CITATION.cff.

The code is open-source and free to use. It is aimed for, but not limited to, academic research. We welcome forking of this repository, pull requests, and any contributions in the spirit of open science and open-source code. For inquiries about collaboration, you may contact Md Shadab Alam (md_shadab_alam@outlook.com) or Pavlo Bazilinskyy (p.bazilinskyy@tue.nl).

Repository layout

Top folder: processing the videos and the analysis

main.py: downloads the videos of the mapping (from the FTP server, else YouTube) and runs YOLO tracking on their segments, writing one detection CSV per segment to the data folders; runs in a loop. helper_script.py holds its download, trimming and tracking code.
analysis.py: the tables, maps and figures of the dataset and the statistics in this README, from the mapping and, when available, the YOLO detections. count_detections.py counts the detections where the CSVs are and saves the counts (analysis_data/detection_counts.csv.gz), so that analysis.py and the GitHub Action can use them without the CSVs.
add_video.py: the page for adding videos and segments to the mapping by hand (templates in templates/).
common.py (configs, secrets, country codes, git, email), custom_logger.py and logmod.py (logging): shared by all scripts.

Top folder: proposing segments from YouTube channels (see Proposing segments)

propose_segments.py: proposes segments, localities, time of day and vehicle of a channel's videos, and serves the review page of a batch.
run_channels.py: works through the channels sheet, batch by batch; refresh_queue.py builds the queue of channels to look at again for new uploads.
review_hub.py: the page listing the batches to review, and applying and pushing reviewed segments.
sync_workers.py: brings the batches of the worker machines to the reviewing machine and the mapping to them.
review_merges.py: merging towns into the city whose metropolitan area they lie in; review_lessons.py: what the reviews say about the proposals.

Folders

utils/: the analysis and figure code used by analysis.py:
- core/parsing.py (the mapping's list cells, the processed length of a segment), core/dataset_stats.py (totals);
- analytics/contributors.py (who added each video), analytics/metrics_cache.py (YOLO class IDs, unique keys);
- plotting/dataset_figures.py (the README figures), plotting/area_figures.py (county, region, smoothed and spike maps), plotting/maps.py (world and continent maps), plotting/map_labels.py (label placement), plotting/hover.py (popups shared by all figures), plotting/io.py (saving figures), plotting/constants.py.
release/: the weekly metadata update and the release files (update_info.py, make_crowd_jsonl.py, split_large_files.py, validation_release.py, weekly_metadata.sh); make_splits.py (the train, validation and test splits and the balancing weights in crowd_splits.csv: grouped by channel or by locality, so no source is in two splits); and the manual audits: sample_audit.py draws the stratified samples, audit_stats.py scores the label audit and evaluate_detections.py the detections from the manual review; annotate.py is the local browser tool for those audits (labels, review of the detector's boxes in frames, tracking errors in clips), which finds the videos and detection files in any number of folders.
tools/: maintenance scripts for the mapping and the data, e.g. download_videos.py, video_inventory.py, update_traffic_index.py, update_params.py and weekly_detection_counts.sh; the review of cuts is detect_cuts.py (candidates), review_cuts.py (the review page) and apply_cuts.py (splits the mapping and the detection files at the confirmed cuts).
dashboards/: pages and widgets that follow the pipeline (its own README).
analysis_data/: data files the analysis reads and writes: country tables (countries.csv, height_data.csv), the YOLO detection counts (detection_counts.csv.gz) and who added which video (video_contributors.csv). The mapping and metadata files (mapping.csv, mapping-NAME.csv, mapping_metadata.csv, crowd_index.jsonl, ...) and cut_times.csv stay in the top folder.
figures/: the figures of this README (PNG, HTML and, when made on purpose, EPS), and figure_hashes.json, which records what each file was made from, so that only changed figures are saved again (see Step 11); readme/ and screenshots/: images of this README and of the figures.
logs/: records kept by the scripts (traffic_index_log.csv, appended by tools/update_traffic_index.py).
archive/: old files no code uses any more (mappingsubmitted.csv).
tests/: tests of the scripts.
.github/workflows/: the weekly refresh of the README statistics and figures (dataset-figures.yml) and the linter.

Run the scripts with the virtual environment active (see Step 11).

Getting started

Tested with Python 3.12.13 and the uv package manager.

The project is configured to automatically select the appropriate PyTorch backend. On systems with a supported NVIDIA GPU and CUDA driver, uv installs a CUDA enabled version of PyTorch. If no supported GPU is detected, it automatically falls back to the CPU version.

Step 1: Install `uv`

uv is a fast Python package and environment manager.

macOS / Linux:

curl -LsSf https://astral.sh/uv/install.sh | sh

Windows PowerShell:

irm https://astral.sh/uv/install.ps1 | iex

Alternative, if Python and pip are already installed:

pip install uv

Step 2: Fix permissions if needed

In some environments, uv may need to create directories under:

macOS / Linux:

~/.local/share/uv/python

Windows:

%LOCALAPPDATA%\uv\python

If these directories were previously created by another user or with elevated privileges, you may encounter an error such as:

error: failed to create directory ... Permission denied (os error 13)

macOS / Linux

mkdir -p ~/.local/share/uv
chown -R "$(id -un)":"$(id -gn)" ~/.local/share/uv
chmod -R u+rwX ~/.local/share/uv

Windows PowerShell

New-Item -ItemType Directory -Force "$env:LOCALAPPDATA\uv"

icacls "$env:LOCALAPPDATA\uv" /grant "$($env:UserName):(OI)(CI)F"

Step 3: Verify `uv`

Check that uv is available:

uv --version

Step 4: Clone the repository

git clone https://github.com/crowd-dataset/crowd.git
cd crowd

Step 5: Install Python 3.12.13

The project is tested with Python 3.12.13.

Install it using uv:

uv python install 3.12.13

The repository contains a .python-version file so that uv can automatically select the correct Python version.

Step 6: Create the virtual environment

Create a virtual environment in .venv using Python 3.12.13:

uv venv --python 3.12.13

Step 7: Activate the virtual environment

macOS / Linux:

source .venv/bin/activate

Windows PowerShell:

.\.venv\Scripts\Activate.ps1

Windows Command Prompt:

.\.venv\Scripts\activate.bat

Step 8: Install the dependencies

Install the project dependencies from pyproject.toml:

uv pip install -r pyproject.toml

The project uses:

[tool.uv.pip]
torch-backend = "auto"

This allows uv to automatically select the appropriate PyTorch backend for the current machine.

For example:

NVIDIA GPU with a compatible CUDA driver → CUDA enabled PyTorch
No supported GPU → CPU PyTorch

No manual selection of a CUDA version should normally be required.

Note: Do not use uv sync for the initial installation if automatic PyTorch backend detection is required. Automatic torch-backend selection currently applies to the uv pip interface.

Step 9: Verify the PyTorch backend

Check which PyTorch backend was installed:

python -c "import torch; print('PyTorch:', torch.__version__); print('CUDA available:', torch.cuda.is_available()); print('CUDA version:', torch.version.cuda); print('Device:', torch.cuda.get_device_name(0) if torch.cuda.is_available() else 'CPU')"

On an NVIDIA CUDA system, the output should look similar to:

PyTorch: 2.8.0+cu129
CUDA available: True
CUDA version: 12.9
Device: NVIDIA GeForce RTX 5090

On a machine without a supported GPU, CUDA available will be False and the CPU version of PyTorch will be used.

Step 10: Create `config` and `secret`

mapping.csv (the videos, segments and localities) is in the top folder of the repository. Copy the two templates in the top folder:

cp default.config config
cp default.secret secret

In config, set data to the folders that hold (or will hold) the YOLO output (bbox/*.csv in each), videos to the folders with the videos, and mapping to the mapping file (mapping.csv); the other parameters are described in Configuration of project. The scripts stop when there is no config. Fill in only the keys of secret you need (The secret file); the analysis needs none.

Step 11: Run the scripts

Run the scripts from the top folder of the repository with the virtual environment active:

python main.py                     # download the videos and run YOLO on their segments (runs in a loop)
python analysis.py                 # tables, figures and README statistics (YOLO counts from the data folders
                                   # or, without them, from analysis_data/detection_counts.csv.gz)
python count_detections.py --data /path/to/data   # count the YOLO detections and save the snapshot

analysis.py saves a figure (each of its HTML, PNG and EPS files) only when what it shows changed since it was last saved, as recorded in figures/figure_hashes.json, and renders the PNG and EPS files in parallel; an unchanged run takes about a minute. To save every figure again, delete figures/figure_hashes.json first.

The scripts in tools/ and release/ can be run as a file or as a module, and from any folder (they switch to the top folder, where mapping.csv and the other files are):

python tools/check_lists.py
python -m tools.check_lists

The tests run with python -m pytest tests (after uv pip install pytest); tests/test_reject_segments.py is a script run with python tests/test_reject_segments.py on the reviewing machine, as it needs its review plans. Some modules have a self-check: python -m utils.core.parsing, python -m utils.plotting.map_labels.

Configuration of project

Configuration of the project needs to be defined in config; copy default.config to start. The scripts stop when config is missing or has fewer entries than default.config. Entries marked optional may be left out. The config file has the following parameters:

Data and mapping

data: folders with the YOLO output (a bbox/ folder of detection CSVs in each).
videos: folders with the downloaded videos.
mapping: the mapping file (mapping.csv).
countries_analyse: ISO3 codes of the countries to analyse; empty for all.
min_confidence: minimum YOLO confidence of a detection counted in the analysis.

Video processing (main.py)

machine_name: name of the machine in the emails sent by main.py.
tracking_mode: run YOLO tracking and write the detection CSVs.
tracking_model: the YOLO model file (e.g. yolo11x.pt).
bbox_tracker: the tracker settings (bbox_custom_tracker.yaml, its track buffer set from track_buffer_sec).
track_buffer_sec: how long (s) a lost track is kept before it ends.
max_workers: segments processed at the same time.
download_workers: videos downloaded or prepared at the same time.
max_active_segments_per_video: segments of the same video processed at the same time: with 1, the workers take segments of different videos (with max_workers=3, 3 videos at once); with 2+, several workers may take segments of one video, which helps when a video has many segments.
runs_root: folder for YOLO's working files.
external_ssd: the videos are on an external drive: segments are trimmed on the internal drive (the last folder of videos).
ftp_server: server from which videos are fetched before trying YouTube (ftp_username and ftp_password, and optionally ftp_token, in secret).
delete_youtube_video, delete_runs_files, delete_labels: delete downloaded videos, YOLO's working files and its label files after processing.
save_annotated_video, save_annoted_img, save_tracked_img, display_frame_tracking: save an annotated video, save annotated or tracked frames, show the tracking while it runs.
snellius_mode, snellius_shard_mode: running as several tasks on the Snellius cluster, each taking its share of the segments.
sleep_sec, git_pull: pause (s) at the end of each loop of main.py, and pulling the repository then.
email_send, email_sender, email_recipients: email when a job finishes or crashes (see the secret file below).

Downloading from YouTube

client: pytubefix client ("WEB", "ANDROID" or "IOS").
need_authentication: sign in to YouTube for downloads with pytubefix.
update_package: update pytubefix and yt-dlp (on Mondays) before downloading.
Optional: use_pytubefix (false: yt-dlp only), cookies_file (a Netscape cookies file for yt-dlp), cookies_browser (yt-dlp's --cookies-from-browser, e.g. chrome) and js_runtime (yt-dlp's JavaScript runtime, e.g. node:/path/to/node).

Updating the mapping

update_pop_country, update_gini_value: update the countries' population and Gini index in the mapping from the World Bank when main.py starts.

Figures and output

font_family, font_size, plotly_template: look of the figures.
logger_level: level of console output: debug, info, warning or error.
Optional: open_html_figures (true: open every saved HTML figure in the browser) and save_eps (false: no EPS figures, as in the GitHub Action; EPS needs poppler's pdftops).

The `secret` file

API keys, passwords and usernames live in a file named secret (no extension) in the root of the project. It is ignored by git, so it is never committed. To create it, copy the template default.secret to secret and fill in the values you need. Any value can stay empty ("") if you do not use the feature it belongs to. Running the analysis on the released dataset needs none of them; running main.py needs at least an empty secret copied from the template.

The file is a single JSON object:

{
  "geonames_username": "",
  "bea_api_key": "",
  "tomtom_api_key": "",
  "restcountries_api_key": "",
  "email_smtp": "",
  "email_account": "",
  "email_password": "",
  "ftp_username": "",
  "ftp_password": "",
  "carto_api_key": "",
  "youtube_api_key": "",
  "locationiq_api_key": ""
}

Which keys you need depends on what you run:

analysis of the released dataset (analysis.py): none; carto_api_key is optional.
adding videos (add_video.py): geonames_username, restcountries_api_key, tomtom_api_key and youtube_api_key; bea_api_key for US localities.
proposing segments from channels (propose_segments.py): youtube_api_key; geonames_username and locationiq_api_key help.
updating the traffic index (tools/update_traffic_index.py): tomtom_api_key.
processing videos (main.py): ftp_username and ftp_password for the VideoFiles server; the email keys only if email_send is on in config.

Key	Used by	What for	Where to get it
`geonames_username`	`add_video.py`, `propose_segments.py`	locality population, a fallback for coordinates, and place lookups	free account at GeoNames; see below
`bea_api_key`	`add_video.py`	GDP of US localities	free key from BEA; see below
`tomtom_api_key`	`add_video.py`, `tools/update_traffic_index.py`	traffic index from TomTom traffic flow	free account at TomTom; see below
`restcountries_api_key`	`add_video.py`	country population, continent and Gini	free account at REST Countries; see below
`email_smtp`, `email_account`, `email_password`	`main.py`	email when a job finishes (`email_send` in `config`)	your email provider; see below
`ftp_username`, `ftp_password`	`main.py`	downloading videos from the VideoFiles server	the maintainers (see Contact)
`carto_api_key`	`utils/plotting/maps.py`	optional: basemap tiles for the Mapbox-style maps without CARTO's watermark; without a key, CARTO's free basemap is used	CARTO; see below
`youtube_api_key`	`propose_segments.py`, `add_video.py`	listing a channel's videos and their length, upload date and channel, without downloading them	see below
`locationiq_api_key`	`propose_segments.py`	optional: faster place lookups (localities from titles, descriptions and chapters, and the towns along a drive "from A to B")	see below

Getting a YouTube Data API key

The key is free; the default quota of 10,000 units a day covers several large channels (listing a channel's search results costs 100 units per 50 videos, video details 1 unit per 50 videos).

Go to the Google Cloud console and sign in with the Google account you want the key to belong to.
Create a project: project selector at the top → New project → give it a name (e.g. crowd) → Create, then select it.
Open APIs & Services → Library, search for YouTube Data API v3 and click Enable.
Open APIs & Services → Credentials → Create credentials → API key. Copy the key that appears.
Recommended: click the new key → under API restrictions choose Restrict key → select YouTube Data API v3 → Save, so the key cannot be used for anything else.
Put the key in secret as "youtube_api_key": "...".

Getting a LocationIQ key (optional)

propose_segments.py looks places up on OpenStreetMap: to guess a video's locality from a named place ("Al Khan Beach" is in Sharjah) and to split a drive "from A to B" into the towns along its route (about 25 lookups per such drive). Without a key it uses the public Nominatim server, which allows one request a second and refuses requests for a while when a busy day goes over that. LocationIQ runs the same OpenStreetMap search with a free key: 5,000 requests a day at two a second, with no payment details. With a key in secret, lookups go to LocationIQ; once its daily requests are used up, they fall back to the public server.

Sign up for the free plan at locationiq.com and confirm your email address.
Open the dashboard's Access Tokens page and copy the token there.
Put it in secret as "locationiq_api_key": "...". Nothing else needs changing; running channel batches pick it up from their next chunk.

Getting a GeoNames username

GeoNames is free; an account allows 10,000 requests a day and 1,000 an hour.

Create an account at geonames.org/login and confirm your email address with the link GeoNames sends.
Sign in and open your account page. At the bottom, click Click here to enable the free web services. Requests are refused until this is done.
Put your username (not your email address) in secret as "geonames_username": "...".

Getting a BEA API key

The U.S. Bureau of Economic Analysis key is free and is used only for localities in the USA.

Fill in your name and email address at apps.bea.gov/API/signup.
BEA emails you the key (a 36-character code); click the activation link in that email.
Put the key in secret as "bea_api_key": "...".

Getting a TomTom API key

The traffic index uses TomTom's Traffic Flow API (segment data). Its free tier allows 20,000 requests a month, with no payment details needed.

Create an account at developer.tomtom.com and confirm your email address.
Open the Dashboard: a key is created with the account (Keys section). Copy it.
Put it in secret as "tomtom_api_key": "...".

tools/update_traffic_index.py uses one request per locality and only queries localities where it is local daytime; requests beyond the free tier are refused unless payment details are added to the TomTom account.

Getting a REST Countries API key

add_video.py uses version 5 of the REST Countries API, which needs a key. The free plan allows 1,000 requests a month, with no payment details.

Sign up at restcountries.com/sign-up and confirm your email address.
Create an API key in the dashboard and copy it.
Put it in secret as "restcountries_api_key": "...".

Email settings

main.py can email you when a job finishes (email_send, email_sender and email_recipients in config). It connects to the SMTP server over SSL on port 465.

email_smtp: the SMTP server of your provider, e.g. smtp.gmail.com for Gmail.
email_account: the full email address you send from.
email_password: for Gmail, an app password, not your normal password. Turn on 2-Step Verification for the Google account, then create an app password at myaccount.google.com/apppasswords and paste the 16 characters. Other providers have similar app passwords or use the account password.

Getting a CARTO key (optional)

The Mapbox-style maps (mapbox_map_*) use CARTO's free basemap, which shows a CARTO watermark. With a key, maps.py loads CARTO's basemap tiles with the key instead.

Create an account at carto.com.
In the CARTO Workspace, create an API access token under Developers → Credentials and copy it.
Put it in secret as "carto_api_key": "...".

Example of YOLO output for a video with dashcam footage

Video: https://www.youtube.com/watch?v=_Wyg213IZDI.

Selection procedure

Verification of segments

After segments were added, they were checked again by hand. The checks were done by Md Shadab Alam, on the main branch, for the segments of mapping.csv, including the videos that other contributors of this repository added (see Who added the videos).

Why a manual step. A retained segment should be one continuous, unedited stretch of driving. Watching all the footage again is not feasible, but an automatic detector of abrupt visual change can point to the few seconds where a cut is likely. The detector alone is not reliable enough to cut the data: a sudden change can be an edit, but also the entrance of a tunnel, a flash of sunlight or a passing lorry, and telling them apart needs context. So the detector proposes candidates, a person decides, and only the decisions of the person are applied. Among the candidates that could be judged so far, most were not cuts (about 59%), so splitting at every candidate would have broken many continuous segments.

1. Candidate cuts. python -m tools.detect_cuts runs FFmpeg scene-change detection (threshold 0.45) on the original video of each segment, over the segment only, merges times closer than 2 s, and writes the candidates to cut_times.csv with the video, the segment start, the frame rate and the time of the candidate. The threshold favours abrupt cuts: gradual transitions, and cuts between similar scenes, can be missed, and the detector does not find speed changes or advertisements.

2. Reviewing. python -m tools.review_cuts serves a page (port 5050) that plays the YouTube video from two seconds before to two seconds after each candidate. The reviewer marks the candidate as yes (a real cut), no (not a cut) or unavailable (YouTube no longer plays or embeds the video), and can add a note. The decision and the time of the review are saved in cut_times.csv.

In total, 52,838 cuts were analysed.

3. What the reviewer looked for.

Cuts in videos added by other contributors. The segments of the videos added by the other authors of the repository were checked for cuts, in the same way as the others.
Videos that are not at normal speed. Videos that play faster or slower than 1x (for example fast-forwarded footage or time-lapses) are not suitable, because speeds and crossing times in them are wrong. Such cases are noted, for example as fast forward.
Advertisements in the footage. An advertisement that was cut into the middle of a video is not footage of the road. Where one was found, it was removed from the segments by the reviewer.

4. Applying the result. python -m tools.apply_cuts applies the candidates marked yes. A cut at time t splits its segment in mapping.csv into a part that ends at floor(t) and a part that starts at ceil(t), so one to two seconds around the cut are dropped, and parts shorter than one second are discarded. The detection CSV of the segment is split at the same frames, with frame numbers restarted at 1 and the true frame rate read from the video (not the rounded one in the file name), so the split files stay aligned. Segments that become too short are removed, and detection CSVs that no longer match a segment in mapping.csv are removed from the data folder.

Description and analysis of dataset

Note: The interactive figures are displayed using htmlpreview.github.io, which may be blocked on some institutional or education networks. If an interactive figure does not open, try another network or download the corresponding .html file from the figures directory and open it locally in a web browser.

Description of dataset

Locations of cities with footage in dataset. Note: continents are based on geography, i.e., the cities in Russia east from Ural mountains are shown as Asia.

Locations of cities with footage in dataset with a density overlay of population. Note: continents are based on geography, i.e., the cities in Russia east from Ural mountains are shown as Asia.

Locations of cities with footage in dataset with a density overlay of the number of videos in the dataset. Note: continents are based on geography, i.e., the cities in Russia east from Ural mountains are shown as Asia.

Locations of cities with footage in dataset with a density overlay of the total number of seconds of footage in the dataset. Note: continents are based on geography, i.e., the cities in Russia east from Ural mountains are shown as Asia.

Amount of footage (hours) per country in Africa. Note: the colour scale is logarithmic; countries are grouped by the continent of their cities in the dataset; a country with cities on two continents (Russia, Türkiye) is coloured and labelled with its whole footage on both maps.

Amount of footage (hours) per country in Asia. Note: the colour scale is logarithmic; countries are grouped by the continent of their cities in the dataset; a country with cities on two continents (Russia, Türkiye) is coloured and labelled with its whole footage on both maps.

Amount of footage (hours) per country in Europe. Note: the colour scale is logarithmic; countries are grouped by the continent of their cities in the dataset; a country with cities on two continents (Russia, Türkiye) is coloured and labelled with its whole footage on both maps.

Amount of footage (hours) per country in North America. Note: the colour scale is logarithmic; countries are grouped by the continent of their cities in the dataset; a country with cities on two continents (Russia, Türkiye) is coloured and labelled with its whole footage on both maps.

Amount of footage (hours) per country in Oceania. Note: the colour scale is logarithmic; countries are grouped by the continent of their cities in the dataset; a country with cities on two continents (Russia, Türkiye) is coloured and labelled with its whole footage on both maps.

Amount of footage (hours) per country in South America. Note: the colour scale is logarithmic; countries are grouped by the continent of their cities in the dataset; a country with cities on two continents (Russia, Türkiye) is coloured and labelled with its whole footage on both maps.

Amount of footage (hours) per US state. Note: the colour scale is logarithmic.

Footage per US county (log colour scale): localities are assigned to the county they lie in. Hover over a county for its localities.

Hours of footage per million people per US county (population: US Census Bureau estimates, 2024).

Share of footage recorded at night per US county (counties with at least an hour of footage).

Footage per European region: Eurostat NUTS 2 regions for the EU, EFTA, the United Kingdom and candidate countries, and geoBoundaries regions for European Russia (west of the Urals), Ukraine, Belarus and Bosnia and Herzegovina; Moldova, Kosovo and small countries as a whole.

Hours of footage per million people per European region (population: Eurostat, each region's latest year; for Russia, Ukraine, Belarus and Bosnia and Herzegovina, Wikidata's latest figure).

Share of footage recorded at night per European region (regions with at least an hour of footage).

One spike per locality in the contiguous United States (Alaska and Hawaii are not in the view); height and colour show its hours of footage (log scale).

One spike per locality in Russia; height and colour show its hours of footage (log scale).

One spike per locality in Japan.

Where the footage is in Europe: hours of footage per 10,000 km², smoothed over about 50 km (log colour scale); grey: no footage nearby. Callouts mark the densest places.

Where the footage is in North America.

Where the footage is in Asia.

Blind spots of the dataset in Europe: distance from every place to the nearest locality with at least an hour of footage. Callouts mark the places furthest from any footage.

Blind spots of the dataset in North America.

Blind spots of the dataset in Asia.

Share of footage recorded at night in Europe, smoothed over about 50 km and weighted by footage; grey where there is less than an hour of footage within 100 km.

Average year of upload of the footage in Europe (weighted by footage): where the footage is recent and where it is older.

Number of videos against the amount of footage per locality (log scales), coloured by continent, with every country and its flag in the legend on the right, in alphabetical order (scroll it in the HTML version; click a country to hide or show it); hover over a point for its details. The 40 localities with the most footage are labelled where there is room, the top 12 in larger black text; in the HTML version, zoom in to label every point.

The same on linear scales, which show how far the localities with the most footage lead, with the same country legend. The 40 localities with the most footage are labelled where there is room, the top 12 in larger black text; in the HTML version, zoom in to label every point.

Number of videos against the amount of footage per country (log scales), coloured by continent; hover over a point for its details. Countries are labelled with their flag and ISO3 code, the 30 with the most footage in larger black text; in the HTML version, zoom in to label every point.

The same on linear scales, which show how far the countries with the most footage lead. The 30 countries with the most footage are labelled with their flag and ISO3 code where there is room; in the HTML version, zoom in to label every point.

Footage (hours) by continent, by day and by night. Note: continents are based on geography, i.e., the cities in Russia east from Ural mountains are shown as Asia.

Number of videos by quarter of upload and continent. The current quarter (hatched) is not complete yet.

Footage by the type of vehicle it is filmed from (log scale), with each type's share of the footage and of its footage at night.

Amount of footage per locality against its population. Large localities with little footage are candidates for adding videos.

Footage per locality against its population (log scales), coloured by whether a locality has more than three times, about, or less than a third of the footage its size predicts (the dotted least-squares line). Labelled: the most populous localities, those with the most footage, and the most under-covered cities of over a million people, which are the next ones to collect.

Amount of footage per locality against its gross metropolitan product (GMP of the metropolitan area: localities in one metropolitan area share its GMP), with the Spearman rank correlation: whether economically large cities dominate the data.

Amount of footage per locality against its traffic index (how much slower than free flow traffic is, from one TomTom reading per locality), for localities with a reading above 0, with the Spearman rank correlation.

Amount of footage per country against country indicators, with the Spearman rank correlation and a trend line per panel, showing whether some kinds of countries are over- or under-represented.

Share of footage recorded at night per country. Countries with less than 10 hours of footage are grey, as their share is not reliable.

Share of footage recorded at night per European country, printed on each (countries with at least 10 hours of footage; Russia and Türkiye whole).

Hours of footage per million inhabitants per country: coverage relative to country size (log scale; countries with less than 10 hours of footage are grey).

Hours of footage per million inhabitants per European country, printed on each (log scale; countries with at least 10 hours of footage; Russia and Türkiye whole).

Hours of footage per million people against GDP per person (both log scales; circle size: population), coloured by whether a country has more than twice, about, or less than half the footage its wealth predicts (the dotted least-squares line). GDP per person at purchasing-power parity is the World Bank's latest value.

Share of each country's footage that comes from its largest YouTube channel. Where it is high, one uploader's routes and camera dominate that country's data. Countries with less than 10 hours of footage are grey.

Type of vehicle the footage is recorded from, as a share of footage per continent.

Share of footage coming from the largest YouTube channels.

Distribution of the length of segments (parts of videos included in the dataset).

Every locality in the dataset as a dot coloured by continent; the dot area is proportional to the hours of footage.

The same localities on a globe. The HTML version spins until you grab it; drag to turn it.

The localities on a 3D satellite globe (NASA Blue Marble), with a spike for each locality whose height shows its hours of footage on a log scale, so localities with little footage are visible too. In the HTML version, a button merges neighbouring localities into one spike per area of about 200 km, which is easier to read where localities are dense. Each spike is orange for the share of its footage recorded by day and blue for the share at night. The HTML version spins until you grab it; drag to turn it, scroll to zoom.

The spinning globe of the CROWD dashboards: every locality as a square that grows with its hours of footage, and the places with segments added in the last week (red) and month (amber). The HTML version loads the latest data when it opens (from the globe on bazilinskyy.github.io, rebuilt every night) and turns until you grab it; drag to turn it, hover over a place for its footage and new segments, and click it to watch its longest segment on YouTube.

Footage by continent, country and locality, coloured by the share recorded at night. A country with localities on two continents (Russia, Türkiye) has a box in each. In the HTML version, click a box to zoom in down to the localities.

Share of footage against share of population (of all countries in the dataset) for the countries among the 25 largest by footage or the 25 largest by population (log scale). Where the orange dot is left of the blue one, the country is under-represented.

Share of footage by population of the locality, per continent and overall: what "urban" means in each part of the dataset.

Effective number of YouTube channels per country (inverse Herfindahl index of the channels' shares of footage): 1 means a single uploader; n means as diverse as n channels with equal shares. Countries with less than 10 hours of footage are grey.

Effective number of YouTube channels against hours of footage per country (log scales; circle size: population), coloured by whether a country has more than twice, about, or less than half the channels its amount of footage predicts. Countries below the line rest on few uploaders, so their footage reflects few channels' routes, cameras and styles.

Where the 20 YouTube channels with the most footage film, numbered by footage, with the number of countries and hours of each (dot area proportional to hours). Travel channels span many countries; most large channels film in one country or city. In the HTML version, each channel's number links to the channel on YouTube.

Length of segments by continent and by type of vehicle (boxes: quartiles; whiskers: 5th and 95th percentiles).

Detections in the footage (YOLO)

Objects detected in the footage by YOLO (yolo11x with tracking): unique tracked persons, cars, bicycles, motorcycles, buses and trucks, counted with a confidence of at least 0.7 and expressed per minute of analysed footage. The counts per segment are kept in analysis_data/detection_counts.csv.gz (video, start of the segment and the unique tracked objects of each class), which python count_detections.py --data <folder with bbox/> writes where the YOLO detection CSV files are (the first run reads about 900 GB of CSVs, in parallel with --workers; later runs read only new or changed files, cached in _output/). analysis.py, also in the GitHub Action, makes the tables and figures below from it, matched to the current mapping, so they stay up to date without the CSV files. Every Sunday at 20:00 (Amsterdam time), tools/weekly_detection_counts.sh on the machine with the data drive counts newly analysed footage and pushes analysis_data/detection_counts.csv.gz if it changed, ahead of the weekly refresh of the figures.

The detections are automatic outputs, not ground truth. What the detector sees depends on the camera, resolution and viewpoint, so the figures by vehicle type and by year of upload show these effects next to the comparisons between countries.

Based on the YOLO detections in 91,983 processed segments, last updated 2026-11-02 00:25 UTC.

Tables

Detections per continent

#	Continent	Analysed (h)	Of footage (%)	Localities	Persons / min	Cars / min	Bicycles / min	Motorcycles / min	Buses / min	Trucks / min
1	Africa	819.6	76.3	563	44.34	47.56	1.32	10.37	2.37	9.90
2	Asia	9,531.9	94.9	2,148	25.86	44.71	1.04	16.52	2.76	8.28
3	Europe	8,767.5	90.2	3,696	17.84	49.18	2.58	1.33	1.94	5.62
4	North America	9,994.7	82.5	2,468	10.05	56.68	0.64	0.42	1.00	8.10
5	Oceania	1,550.6	80.8	437	8.72	60.13	0.58	0.62	1.35	9.04
6	South America	1,175.3	88.7	398	19.70	53.06	1.33	4.65	1.80	7.21
7	All	31,839.5	87.9	9,710	18.10	50.83	1.33	5.91	1.87	7.53

Top 20 countries by pedestrians per minute (at least 10 analysed hours in at least 3 localities)

#	Country	Pedestrians / min	95% interval	Analysed (h)	Localities
1	🇺🇬 Uganda	98.48	85.51–114.36	84.9	19
2	🇳🇵 Nepal	98.26	27.58–102.13	38.2	4
3	🇧🇩 Bangladesh	66.25	45.15–72.26	107.2	28
4	🇳🇬 Nigeria	66.14	39.72–78.00	66.3	35
5	🇰🇪 Kenya	58.79	54.46–65.94	19.9	55
6	🇨🇺 Cuba	56.92	30.19–61.26	34.3	25
7	🇵🇰 Pakistan	55.79	42.01–76.81	146.7	37
8	🇪🇹 Ethiopia	48.57	47.40–81.71	42.2	8
9	🇮🇳 India	47.50	43.12–54.57	476.8	291
10	🇻🇳 Vietnam	45.71	27.37–48.44	1,115.7	68
11	🇰🇭 Cambodia	45.55	25.44–47.25	221.8	7
12	🇬🇲 Gambia	43.04	35.10–55.98	95.4	29
13	🇨🇬 Congo (Congo-Brazzaville)	42.32	5.83–43.13	11.8	3
14	🇲🇼 Malawi	41.64	32.74–55.76	31.5	49
15	🇸🇱 Sierra Leone	39.48	24.98–45.91	79.1	46
16	🇨🇴 Colombia	38.42	34.78–42.84	50.9	36
17	🇬🇭 Ghana	35.95	30.04–59.72	11.4	10
18	🇮🇩 Indonesia	34.94	27.79–45.60	1,018.6	176
19	🇲🇦 Morocco	32.84	28.53–39.51	47.3	35
20	🇬🇧 United Kingdom	31.64	9.89–35.38	1,510.1	358

Top 20 localities by pedestrians per minute (at least 1 analysed hour)

#	Locality	Pedestrians / min	Analysed (h)
1	🇮🇳 Varanasi	277.62	1.0
2	🇨🇫 Bangui	123.28	5.2
3	🇮🇳 Anantapur	118.73	1.2
4	🇺🇬 Mubende	109.17	3.2
5	🇲🇬 Toamasina	108.70	1.3
6	🇵🇰 Dera Ismail Khan	108.37	1.2
7	🇵🇰 Sialkot	105.14	3.4
8	🇺🇬 Kabale	103.17	3.1
9	🇳🇵 Kathmandu	102.68	35.7
10	🇸🇱 Kenema	100.05	2.4
11	🇵🇰 Multan	99.96	9.8
12	🇺🇬 Kampala	98.79	68.2
13	🇲🇸 Salem	98.57	1.2
14	🇵🇰 Mansehra	93.53	2.1
15	🇮🇳 Gaya	91.63	3.0
16	🇸🇱 Bo	90.46	2.6
17	🇸🇱 Makeni	88.17	1.2
18	🇲🇬 Antananarivo	85.38	2.7
19	🇷🇼 Kigali	83.35	2.5
20	🇳🇬 Lagos	82.00	34.3

Figures

Share of each country's footage that has been analysed with YOLO (countries with less than 10 hours of footage are grey).

Pedestrians per minute of footage per country (unique tracked persons; countries with less than 10 analysed hours are grey).

Pedestrians per minute of footage per European country (countries with at least 10 analysed hours; Russia and Türkiye whole).

Pedestrians per minute of footage per country, with a 95% bootstrap interval over its localities (countries with at least 10 analysed hours in at least 3 localities).

How much footage a stable estimate of a locality needs: the width of its 95% bootstrap interval of pedestrians per minute over its segments (as a share of the rate) against its analysed hours (localities with at least 5 segments; segments of one drive are alike, so the true intervals are somewhat wider).

Pedestrians against cars per minute per locality (at least an analysed hour), split at the medians into walkable, busy, car-dominated and quiet streets.

Pedestrians per minute by population of the locality, per continent, with a 95% bootstrap interval over localities (groups with at least 3 localities and an analysed hour).

Road-user mix of the 30 localities with the most analysed footage, each as 100 squares (ordered by the share of pedestrians).

Detected road users per minute of footage, per class, in the 40 countries with the most analysed footage (log colour scale).

Pedestrians per minute of footage per locality, by continent.

Distribution of pedestrians per minute over segments of at least a minute (log scale).

Mix of detected road users per continent.

Mix of detected road users in the 30 countries with the most analysed footage.

Share of two-wheelers (motorcycles and bicycles) among the detected vehicles per country (countries with less than 10 analysed hours are grey).

Detected road users per minute by day and by night, per class and continent (log scale).

Pedestrians per minute of footage by day and by night per country (countries with at least an hour of each).

Pedestrians per minute by the type of vehicle the footage is filmed from: the viewpoint changes what the detector sees.

Detected road users per minute by year of upload (years with at least 10 analysed hours): newer cameras and higher resolutions can change what the detector sees.

Pedestrians per minute of footage per country (at least 10 analysed hours) against country indicators: population, road deaths, Gini index, median age and literacy (from the mapping), GDP per person, urban population, density and jobs in agriculture (World Bank), pedestrians among road deaths, registered vehicles, physical inactivity and obesity (WHO), mean temperature and rainfall (ERA5, World Bank climate portal) and the footage-weighted TomTom traffic index. Dashed lines: medians of equal-sized groups of countries along each indicator.

Pedestrians per minute of footage against GDP per person (countries with at least 10 analysed hours; circle size: population), coloured by whether a country has more than 1.5 times, about, or less than two-thirds of the pedestrians its wealth predicts.

Motorcycles as a share of detected vehicles against GDP per person, coloured by whether a country has more than twice, about, or less than half the share its wealth predicts.

Bicycles per minute against GDP per person: cycling cultures stand out above the line.

Cars per minute against GDP per person: motorisation on the street, relative to wealth.

Road traffic deaths per 100,000 people against pedestrians per minute: countries with more or fewer deaths than their number of pedestrians on the street predicts.

Pedestrian road deaths per 100,000 people (road deaths times the WHO's pedestrian share of them, 2016) against pedestrians per minute: where walking is more or less deadly than the number of pedestrians on the street predicts (not countries where most road deaths are of an unknown road user, e.g. the Philippines).

Pedestrians per minute per US county (counties with at least an hour of analysed footage).

Pedestrians per minute per European region (NUTS 2, and first-level regions of European Russia, Ukraine, Belarus and Bosnia and Herzegovina; regions with at least an hour of analysed footage).

Bicycles per minute per European region: cycling regions stand out.

Pedestrians per minute in Europe, smoothed over about 50 km; grey where there is less than an analysed hour within 100 km.

Walkable street minutes in Europe: the share of analysed minutes with at least 2 pedestrians and at most 10 cars per minute, smoothed over about 50 km.

Walkable street minutes per European country: the share of analysed minutes with at least 2 pedestrians and at most 10 cars per minute (countries with at least 10 analysed hours; Russia and Türkiye whole; the colours end at the 95th percentile, higher shares get the darkest).

Motorcycles as a share of detected vehicles in Europe, smoothed over about 50 km.

Bicycles per minute in Europe, smoothed over about 50 km.

Pedestrians per minute per locality against its population (localities with at least an hour of analysed footage): busy and quiet streets for a city's size.

Pedestrians per minute of footage per locality against its population.

Adding videos to dataset

To add videos to the mapping file by hand, run python add_video.py. It opens a web form (a Flask app on a local port) for one video at a time. Fill in the locality, the state (optional, recommended for the USA 🇺🇸 and Canada 🇨🇦), the country (mandatory, chosen from a list) and the YouTube link, and click Fetch data. While you type the locality, the form suggests localities that are already in the mapping file.

Adding a video to Delft, Netherlands 🇳🇱 (with no state), before fetching its data. While the locality is typed, the form suggests the matching locality already in the mapping file.

Fetch data checks whether the locality is already in the mapping file and looks up the rest. For the video: its upload date and channel from YouTube. For a new locality: its coordinates, population and gross metropolitan product (GDP), the country's population, road deaths, continent, literacy rate, average height, median age and Gini index, and the traffic index (see The secret file for the services used). The video plays next to the form, with the current second of the player, the last second of the video and a map of the locality.

The same video after fetching its data, from before Delft was in the mapping file: with no entry for Delft yet, the form adds it as a new locality. Its coordinates are shown on the map under the video; dragging the marker adjusts them.

Kyiv, Ukraine 💙💛 is in the mapping file (with 569 videos), so the form adds the video to it. This video is in the mapping file already: the form shows its existing segments in red (here from second 1 to 1408) next to the player, so a new segment is added after them or the existing ones can be checked.

For each segment (part of a video), add the Time of day (day or night), the Vehicle the footage is filmed from, the Start time (seconds) and the End time (seconds) (larger than the start), and click Submit. A video can have several segments, each starting at or after the end of the previous one; the values of the video (upload date, channel) and of the locality are kept, so for each further segment only its start, end, time of day and vehicle need to be filled in. All values can be corrected in the form before submitting, and by hand in the mapping file later.

Shortcuts and click events

The form accepts the following shortcuts (when the cursor is not in a text field; they also work with non-Latin keyboard layouts):

A: the current second of the video into "Start time (seconds)".
S: the current second of the video into "End time (seconds)".
D: the whole video: "Start time (seconds)" 0 and "End time (seconds)" the last second of the video. Clicking on "Last second" under the video does the same (and copies it).
Q: "Day" for "Time of day".
W: "Night" for "Time of day".

Proposing segments from YouTube channels

Instead of finding segments by hand in add_video.py, whole YouTube channels can be processed automatically: propose_segments.py downloads each video of a channel at low resolution, proposes segments, localities, time of day and vehicle type, and a curator reviews the proposals on a web page and applies the approved ones to the mapping file. Nothing is written to the mapping before a curator applies it.

1. Channels to process. run_channels.py takes channels from a Google Sheet (the sheet URL is set in run_channels.py): column A is the channel URL, column B its status and column C its country. It works from the bottom of the sheet up, through channels with an empty status, and proposes segments for each in chunks of videos (--chunk, default 15), each chunk in a new process, so changes to the code apply from the next chunk:

python run_channels.py

It runs until stopped (--max-minutes N stops after N minutes) and continues where it stopped when restarted. It cannot write to the sheet: it logs lines starting with SHEET: (set row N to Processing, Processed or Rejected) for the curator to copy into the sheet. Without a country in column C, it uses the channel's country on YouTube. A single channel can also be processed directly:

python propose_segments.py --channel https://www.youtube.com/@CHANNEL --country "Netherlands" --limit 20

The country must be written as in the add_video.py form. --single-city treats all of a channel's videos as one city (no drives from A to B), and --no-download proposes whole videos from YouTube metadata only. Downloading needs yt-dlp, ffmpeg and node; when YouTube asks to sign in, put a cookies.txt exported from your browser in the root of the project. The youtube_api_key in the secret file is needed to list channels (see The secret file).

2. What is proposed automatically. Videos are left out from their metadata when they are shorter than 5 minutes, live, portrait, or walking tours, time-lapses, compilations, crash videos, highway drives or road trips. For the others, a 160x90 copy is analysed at 2 frames per second:

segments: stationary stretches at the start and end are trimmed and stops longer than 3 minutes are cut out; hard cuts (edits), blank frames and title cards end a segment; footage with the camera facing the driver, not facing forward or shaking, and rural or highway stretches (from YOLO detections) are left out; segments shorter than 30 seconds are dropped. Whole videos are left out for picture-in-picture or skipped footage, and a channel where ten videos in a row have skips is rejected;
time of day: night when the sky or the street is dark for at least a minute, with a hint at dusk;
vehicle: the most common vehicle type of the channel's videos already in the mapping (a car, with a note, for a new channel);
locality: from the title, then the description, then the chapters (a video with chapters for several places is split), and otherwise a named place looked up on OpenStreetMap; drives "from A to B" are split into the towns along the route; a town within 30 km of a city with over a million people counts as that city unless it has 250,000 people itself;
re-uploads of a video already in the mapping are recognised and left out.

The proposals of a channel are kept in _output/proposals/<channel>/plan.json, with frames and signals of each video. Videos left out are listed on the review page with the reason:

Videos of the same channel as the Kyiv example of add_video.py, left out automatically: a video under 5 minutes, a walk, and a drive that skips footage (an edit cut).

3. Reviewing. Run python review_hub.py and open http://127.0.0.1:8770: it lists the channels with videos to review, and clicking one opens its review page (also available directly with python propose_segments.py --review _output/proposals/<channel>/plan.json).

The review hub: channels with videos still to decide. For each video, the page shows the player, the title and description, a contact sheet of frames, a route map, and the proposed segments with the reasons for each cut. Per segment, the curator can change the start and end, switch day and night, approve, reject, split it at a time, move it to another locality, or add a new segment; per video, change the vehicle, locality, state and country (with a map for a new locality), approve the rest of the video or reject it. Shortcuts: A sets the start and S the end of a segment to the current time of the player, D takes the whole video, Q and W set day and night, and F puts the current time in the split box.

The proposal for the Kyiv video used in the add_video.py example above: one segment from 0:01 to 23:30 in Kyiv, Ukraine (the locality guessed from the title, the vehicle from the channel's other videos in the mapping), with a contact sheet of frames and the video's description; it matches the segment added by hand (seconds 1 to 1408).

4. Applying. Apply adds the approved segments to the mapping file through the same code as the add_video.py form: the video, its segments, time of day, vehicle type, upload date and channel, and for a new locality a new row with its coordinates, population and other values. Applied segments are marked in plan.json and cannot be changed on the page any more. Overlapping segments are refused, and writes to the mapping file are locked, so several review pages and a running channel batch can be used at once. Do not stop a review page while it is applying.

5. Merging towns into cities. python review_merges.py serves a page at http://127.0.0.1:8772 with towns of under 250,000 people within 30 km of a city of over a million in the same state, side by side on maps: Y merges a town into the city, N keeps it separate. python review_merges.py --apply merges the accepted towns: their videos move to the city and their name is added to the city's other names.

Weekly update of video metadata

mapping_metadata.csv holds the YouTube metadata of every video in the mapping file (title, upload date, channel, views, description, chapters), and crowd_index.jsonl is the same dataset as JSON Lines, one object per segment, built from mapping.csv and mapping_metadata.csv without the free-text fields (titles, descriptions, chapters). Both are refreshed every Monday at 02:00 (Amsterdam time) by release/weekly_metadata.sh, run by cron on a Mac mini at TU/e in a fresh shallow clone made for each run and removed after it:

it resets the clone to the latest main on GitHub;
python -m release.update_info fetches the metadata of videos added since the last run and retries the videos with missing values (offline videos in the mapping keep being retried). Each video is written to the file as soon as it is fetched (a new video is appended as a row, so an interrupted run keeps what it has; a failed retry leaves its row as it was); rows of videos no longer in the mapping are dropped;
python -m release.make_crowd_jsonl rebuilds crowd_index.jsonl;
python -m release.split_large_files split splits a file over 99 MB into NAME_part1.EXT, NAME_part2.EXT, ... (each under 99 MB; CSV parts repeat the header), as GitHub refuses files over 100 MB;
the two files (or their parts) are committed as "Weekly update of video metadata" and pushed. A file over 99 MB is never pushed: the job stops instead. When someone pushed in the meantime, the commit is replayed on top of theirs.

When the files are split, the whole file is not in the repository. release/update_info.py, release/make_crowd_jsonl.py and release/validation_release.py rebuild it from its parts when needed, or run python -m release.split_large_files join mapping_metadata.csv crowd_index.jsonl. The log of the job is in ~/Library/Logs/crowd-metadata.log on the mini. To run it elsewhere, clone the repository with push access and add 0 2 * * 1 /path/to/clone/release/weekly_metadata.sh >> /path/to/crowd-metadata.log 2>&1 to crontab -e (set PY to the Python of the environment if it is not ~/micromamba/envs/crowd/bin/python).

Who added the videos

Videos are added by several people, each in their own copy of the mapping file: mapping.csv is Pavlo's, mapping-olena.csv Olena's, mapping-shadab.csv and mapping-epfl.csv Shadab's and, in general, mapping-NAME.csv NAME's. Videos added to mapping.csv in commits from a contributor's own account (currently Shadab's and Faye's) are credited to that contributor. Each video is credited to the file it first appeared in. The credits are kept in analysis_data/video_contributors.csv: one row per video with the contributor, the time of the commit that added it (UTC) and the mapping file it first appeared in. It was built from the git history with python -m utils.analytics.contributors, and analysis.py (also in the GitHub Action) adds new videos from the commits of the last 60 days, so a video Olena added to mapping-olena.csv before it was reviewed into mapping.csv is hers; videos in no commit yet get the time of the last commit to their mapping file. It is the only record of who added what once the history of the mapping files is gone, so keep it when rewriting the repository's history. A video that appeared in mapping.csv and another file at the same time is credited to mapping.csv, as the other files are refreshed with copies of it. The table counts the videos in the dataset now.

#	Contributor	Videos	Footage (h)	Localities	Countries	First added	Last added
1	Pavlo	74,649	32,938.3	7,819	234	2025-01-19	2026-11-01
2	Olena	4,883	1,762.6	951	91	2025-05-03	2026-10-29
3	Shadab	2,818	1,445.7	500	96	2023-12-17	2026-07-22
4	Margarida	54	24.9	34	4	2025-05-02	2025-06-05
5	Aloysia	38	18.0	17	2	2025-04-29	2025-07-08
6	Faye	42	14.8	29	11	2026-06-30	2026-07-12

Footage in the dataset over time, by who added it (cumulative hours of the videos in the dataset now, by the date they were added).

Contact

If you have any questions or suggestions, feel free to reach out to Md Shadab Alam (md_shadab_alam@outlook.com) or Pavlo Bazilinskyy (p.bazilinskyy@tue.nl).

Licence

Code: MIT License
CROWD dataset: Creative Commons Attribution 4.0 International (CC BY 4.0)

The dataset licence covers CROWD generated structured data, annotations, and derived computer vision outputs. It does not cover the underlying YouTube videos or other third party content.

CROWD does not redistribute the original YouTube videos, extracted frames, or images. Source videos remain subject to the rights of their respective owners and YouTube's Terms of Service.

Name		Name	Last commit message	Last commit date
.github		.github
analysis_data		analysis_data
archive		archive
dashboards		dashboards
figures		figures
logs		logs
readme		readme
release		release
screenshots		screenshots
templates		templates
tests		tests
tools		tools
utils		utils
.gitignore		.gitignore
.jscpd.json		.jscpd.json
.python-version		.python-version
CITATION.cff		CITATION.cff
DATASHEET.md		DATASHEET.md
DATA_LICENSE.md		DATA_LICENSE.md
LICENSE		LICENSE
README.md		README.md
SECURITY.md		SECURITY.md
__init__.py		__init__.py
add_video.py		add_video.py
analysis.py		analysis.py
bbox_custom_tracker.yaml		bbox_custom_tracker.yaml
common.py		common.py
common_words.txt		common_words.txt
count_detections.py		count_detections.py
crowd_index.jsonl		crowd_index.jsonl
crowd_splits.csv		crowd_splits.csv
custom_logger.py		custom_logger.py
cut_times.csv		cut_times.csv
default.config		default.config
default.secret		default.secret
helper_script.py		helper_script.py
logmod.py		logmod.py
main.py		main.py
mapping-olena.csv		mapping-olena.csv
mapping.csv		mapping.csv
mapping_metadata.csv		mapping_metadata.csv
mapping_remaining.csv		mapping_remaining.csv
plate_codes.py		plate_codes.py
plate_codes_de.py		plate_codes_de.py
propose_segments.py		propose_segments.py
pyproject.toml		pyproject.toml
refresh_queue.py		refresh_queue.py
review_hub.py		review_hub.py
review_lessons.py		review_lessons.py
review_merges.py		review_merges.py
run_channels.py		run_channels.py
sync_workers.py		sync_workers.py
track_ids.py		track_ids.py
uv.lock		uv.lock

Latest commit

History

Folders and files

Repository files navigation

City Road Observations With Dashcams (CROWD)

Dataset overview

Citation and usage of code

Repository layout

Getting started

Step 1: Install uv

Step 2: Fix permissions if needed

macOS / Linux

Windows PowerShell

Step 3: Verify uv

Step 4: Clone the repository

Step 5: Install Python 3.12.13

Step 6: Create the virtual environment

Step 7: Activate the virtual environment

Step 8: Install the dependencies

Step 9: Verify the PyTorch backend

Step 10: Create config and secret

Step 11: Run the scripts

Configuration of project

The secret file

Getting a YouTube Data API key

Getting a LocationIQ key (optional)

Getting a GeoNames username

Getting a BEA API key

Getting a TomTom API key

Getting a REST Countries API key

Email settings

Getting a CARTO key (optional)

Example of YOLO output for a video with dashcam footage

Selection procedure

Segments of videos were not selected if frames are skipped

Snippets of videos are not analysed during the movement of camera

Videos are excluded from analysis if the camera is unstable or shaking

Snippets of videos captured in parking areas were excluded from analysis

Videos are excluded from analysis if another video is a part of main video

Verification of segments

Description and analysis of dataset

Description of dataset

Detections in the footage (YOLO)

Tables

Figures

Adding videos to dataset

Shortcuts and click events

Proposing segments from YouTube channels

Weekly update of video metadata

Who added the videos

Contact

Licence

About

Topics

Resources

Security policy

Stars

Watchers

Forks

Releases

Packages

Used by

Contributors

Languages

Step 1: Install `uv`

Step 3: Verify `uv`

Step 10: Create `config` and `secret`

The `secret` file