Analyzing the Impact of Depth Features on Point Track Performance

Abstract

The multi-object tracking and segmentation task in urban traffic scenes in for improving autonomous driving, poses ongoing challenges from occlusions, lighting variations, and background noise interferences. We tackle the issue of identity switches by enhancing the existing PointTrack framework, by incorporating raw and monocularly estimated depth information into the color-offset tracking pipeline. By combining depth cues directly into the offset features, our approach strengthens geometric reasoning and leads to improved object association in cases of occlusions and reappearances. On the KITTI multi-object tracking and segmentation dataset, our method reduces identity switching by 21.11% compared to PointTrack baseline, showing increased robustness of tractlet association in challenging scenes. Overall, the approach evaluated notably reduces the occurrence of excessive ID switches, which are a major handicap in real, complicated settings. Numerically, our model performs better by having fewer ID switches while maintaining and in certain cases, enhancing the overall MOTSA score.