VisionCart Perception Suite Manifest

Project Identity

Engineering name: VisionCart Perception Suite
Domain: embedded visual perception for a robot car
Hardware target: OpenART Plus / OpenMV for the ROI recognition module
Main deliverables: maze mapping module, ROI segmentation and classification module, OpenMV deployment scripts, evaluation reports

module	role	key files
`maze-mapping`	convert camera observations into stable grid-map elements and control payloads	`main.py`, `origin.py`, `test_lab_thresholds.py`, `test_lab_vote.py`
`roi-segmentation-classification`	train, evaluate, and deploy low-resolution target recognition	`tools/train/`, `tools/eval/`, `deploy/openmv/`, `reports/`
`artifacts`	local-only datasets and model binaries	`datasets/README.md`, `weights/README.md`

Python implementation and evaluation scripts: about 5.9k lines across maze mapping, training, evaluation, and OpenMV deployment.
Raw image/video material: thousands of image and mask files plus sampled video evaluation.
Final video test: 14,601 sampled frames with stability-interval analysis.
Model variants: Keras training checkpoints, float32 TFLite exports, int8 OpenMV exports.
Deployment handling: segmentation-model path, pure-algorithm fallback path, memory release, SD-card-safe execution.

Upload dataset archive and paste URL in README.md and artifacts/datasets/README.md.
Upload weight archive and paste URL in README.md and artifacts/weights/README.md.
Keep artifacts/datasets/** and artifacts/weights/** ignored by git.
Keep reports, demo images, README files, scripts, labels, metrics, and manifests tracked.
Confirm OpenMV SD package can be reconstructed from deploy/openmv/ plus artifacts/weights/openmv/.