Using camera tracking to improve 3D motion sensing in autonomous vehicles

LiDAR scene flow estimation—tracking 3D motion for each point—is critical for autonomous driving but struggles to separate static from dynamic points when data is sparse or occluded. TrackCue addresses this by leveraging dense image-space trajectories from point tracking to provide motion cues that geometric observations alone cannot capture. The method compensates for ego-motion in the image plane, isolates true object motion, and lifts these visual motion cues back to LiDAR points to refine static-dynamic labels. Experimental results show significant improvements in dynamic label precision and F1 score, with corresponding gains in self-supervised scene flow learning.