LiftDepth Risk: Road-Aware Monocular BEV Occupancy and Object Risk
CompletedA camera-only perception pipeline that turns KITTI driving video into metric depth, road-aware BEV occupancy, tracked object footprints, and object-aware risk heatmaps.
About this project
LiftDepth Risk is a camera-only perception demo that converts KITTI driving video into road-aware BEV occupancy and object-level risk maps.
The pipeline combines Depth Anything V2 for metric depth, SegFormer for road/sidewalk segmentation, and YOLO for road-user tracking. Detected vehicles and pedestrians are filtered using the road mask, projected into bird’s-eye view with KITTI camera intrinsics, and represented as compact object footprints instead of noisy dense-depth blobs.
The final output is an interactive HTML demo showing synchronized RGB tracks, depth, segmentation, projected road BEV, object occupancy, and risk heatmaps. The goal is to make monocular scene understanding easier to inspect for robotics and autonomous-driving perception workflows.