[EXTERNAL] docs (vision-track): align subject and audit with measurable validation criteria

[vpollo11]

Changes ✅ :

• Added new “Performance & Validation” section to the VisionTrack subject.

  • Defined measurable success thresholds for detection precision, recall, F1-score, and FPS.
  • Specified required output artifacts:
    • models/checkpoints/best.pt and best_quantized.onnx
    • reports/performance_metrics.json with precision, recall, F1, FPS, and latency
    • reports/demo_results/ with ROI counting screenshots and videos
    • logs/app_errors.log for runtime errors
  • Added GPU/CPU fallback test requirement and ROI counting validation.
  • Included expected output directory summary for consistency with audit.

• Updated the VisionTrack audit for full 1:1 alignment with the subject.

  • Added explicit file existence checks (app.py, counting_logic.py, helpers, YOLO model).
  • Added dependency import test for torch, supervision, cv2, and streamlit.
  • Introduced GPU detection and CPU fallback checks.
  • Added validation for saved YOLO model weights and quantized versions.
  • Required validation of ROI counting overlay and dynamic updates.
  • Added measurable performance thresholds:
    • Precision ≥ 0.85
    • Recall ≥ 0.80
    • F1 ≥ 0.85
    • FPS ≥ 15 (720p)
  • Introduced explicit fail conditions for metrics, FPS, crashes, and missing files.
  • Added final checks to ensure end-to-end pipeline validation:
    data → detection → tracking → counting → dashboard.

• Result: VisionTrack subject and audit are now fully aligned, reproducible, and technically verifiable.

[Oumaimafisaoui]

LGTM