[MatthiasHertelArm]

The traffic counter as developed on MatthiasHertelArm/cmsis-executorch-traffic-counter, imported in one commit (the full history stays there), plus four commits for review:

• The README describes the two prompts to an AI coding agent that created and then optimized the example.
• A third target-type, SSE-320-sim: the Corstone-320 (Cortex-M85, Ethos-U85-256) on the Corstone-320 simulator v0.2.0, which runs natively on Linux, macOS and Windows and serves the CMSIS Debugger, so the test-image run needs neither Docker nor an FVP license. It links a second export of the model for the Ethos-U85 (ai_layer_u85/). CI builds it and runs the Linux release headless.
• The names follow the Ultralytics brand guidelines: "Ultralytics YOLO26n" on first mention, a trademark and no-affiliation note next to the licences, and the panel banner writes YOLO26n as the product is named (the 5x7 font gains lower-case glyphs for that).

What it is

One CMSIS solution, traffic/traffic.cproject.yml, on two target-types with the same Cortex-M55 and Ethos-U55-256:

Target-type	Hardware	Runs
AppKit-E7	Alif Ensemble E7 AppKit (Gen 2), M55 HP core	camera, panel, joystick, SDS recording and playback over the User USB
SSE-300-U55	Corstone-300 FVP	the detector and the tracker on the test image, Test_result: PASS
SSE-320-sim	Corstone-320 simulator (Cortex-M85, Ethos-U85-256)	the same, natively, with the same two trucks (0.65, 0.50)

Ultralytics YOLO26n without its attention blocks, fine-tuned on the COCO images with vehicles, exported with ExecuTorch and compiled by Vela, runs whole on the NPU: 31 ms per 416x416 frame, 25 fps with the camera. A tracker on the Cortex-M55 counts each vehicle once, per class and direction, as it crosses a line on the picture. The AI layer, the weights and the pack lock file are committed so that a clone builds the validated image. CI exports the AI layer from a fresh clone, builds both targets and runs the FVP.

Verification

• AppKit-E7 image builds clean with Arm Compiler 6 (one pre-existing warning in a scatter file comment).
• The brand commit changes the panel banner, so the image is no longer byte-identical to the one validated on the board; the new glyphs were checked by rendering the font on the host, not yet on the panel.
• Licences: Apache-2.0 for the code, the YOLO26n weights under Ultralytics' AGPL-3.0, ExecuTorch sources under BSD-3-Clause. The test image is generated from a COCO photo at build time and not committed.