Robust 3D Multi-Object Tracking in Adverse Weather with Hard Sample Mining

3D multi-object tracking (MOT) is essential for numerous applications such as autonomous driving and robotics. However, the performance of existing 3D MOT solutions can be severely degraded in adverse weather with the risk of missed detection and wrong association. In this paper, we assume that the objects that could be missed or wrong associated with other objects are more meaningful for performance improvement. Based on this facts, an adaptive hard sample mining algorithm is integrated into a two-branch architecture to improve the robustness of 3D MOT in adverse weather. Specifically, we propose a two-branch architecture to learn the region proposals from point clouds and RGB images, respectively. To reduce the risk of missed detection and wrong association, we introduce the hard sample mining to enhance the performance for region proposals. Meanwhile, we dynamically adjust the weights of hard samples during training to achieve an optimal balance between object detection and embedding feature extraction. Our proposed solution is evaluated both on the KITTI tracking dataset and a synthesized foggy dataset. Experimental results show that our proposed solution shows competitive performance in both clear weather and degrading weather. This implies that our solution is able to mitigate the missed detection and reduce the wrong association with good generalization performance.