lane-detection

23:47 · askedSurvey these directories under /Users/liwimoi/projects for computer-vision / perception capabilities relevant to dashcam accident analysis: vlm2, new-knia, hd-map, lane-detection, CarAccident, knia-dag, codex-knia, vlm, labeling, and /Users/liwimoi/aihub/095.교통사고 영상 데이터 (only its top-level scripts and the location_identified_videos*/geoloc_candidates/location_index dirs — do not open .venv, node_modules, videos, or frames). Skip /Users/liwimoi/projects/_archive entirely. Read README.md, CLAUDE.md, docs/*.md, and module/file names; open source files only to confirm what a module does (skim headers/docstrings), not to read everything. Thoroughness: very thorough on the directory sweep, light on file contents. For each project report: (1) one-line purpose; (2) every perception module found — object detection (model/weights), multi-object tracking, re-identification, optical flow / ego-motion, collision/contact detection, lane-line detection, lane-index assignment, road-marking / safety-zone segmentation, traffic light / sign detection, turn-indicator (blinker) detection, brake-light detection, OSD/OCR (timestamp, speed overlay), camera viewpoint (front/rear) detection, geolocation / map matching / HD-map use, BEV / IPM, depth or distance estimation; (3) for each module: inputs → outputs, model and version, any measured accuracy or evaluation numbers stated in docs, and whether it appears runnable (has an entrypoint/venv/weights) or is a stub/plan; (4) known limitations stated in docs. Also note any data assets (HD map coverage areas, geolocated clip subsets and their counts) that a CV step could use. Report as a compact table per project plus a final consolidated table: capability → best existing implementation (project/file) → maturity → measured number if any. Do not pr…