Overview
Goal was to track pedestrians on the ground plane (BEV 2D) from a moving vehicle using multi-sensor inputs (camera, LiDAR) by producing produce stable pedestrian trajectories over time despite noisy detections, occlusions, and ego-motion.

Contributions

• Designed a ground-plane projection method that uses the bottom-midpoint of 2D boxes, calibrated extrinsics, and ego pose to obtain reliable pedestrian footpoints in BEV.

• Implemented and tuned a constant-velocity Kalman filter with dataset-matched process and measurement covariances for stable trajectory estimation.

• Added a pragmatic pre/post-processing stack with NMS, ~35 m range cutoff, confidence thresholds, and outlier filtering, and evaluated with HOTA to diagnose that detector quality is the primary performance bottleneck.

Result
Achieved stable BEV trajectories on cleaner sequences with consistently high association quality and minimal ID switches.

Languages and tools used: Python, Open3D