BEV Object Tracking for LIDAR-based Ground Truth Generation

Building ADAS (Advanced Driver Assistance Systems) or AD (Autonomous Driving) vehicles implies the acquisition of large volumes of data and a costly annotation process to create labeled metadata. Labels are then used for either ground truth composition (for test and validation of algorithms) or to set-up training datasets for machine learning processes. In this paper we present a 3D object tracking mechanism that operates on detections from point cloud sequences. It works in two steps: first an online phase which runs a Branch and Bound algorithm (BBA) to solve the association between detections and tracks, and a second filtering step which adds the required temporal smoothness. Results on KITTI dataset show the produced tracks are accurate and robust against noisy and missing detections, as produced by state-of-the-art deep learning detectors.