Centralized Motion-Aware Enhancement for Single Object Tracking on Point Clouds

Abstract

3D Single Object Tracking (SOT) in LiDAR point clouds has broad application prospects in computer vision, and objects are usually represented by 3D boxes in point clouds. Current methods mostly follow the representation matching-based siamese pattern. However, due to the severe sparse, incomplete shapes of LiDAR point clouds, and the fact that objects in the 3D world do not follow any specific orientation, these are common obstacles to point cloud tracking. In this paper, we propose to represent 3D objects as points, using a key point detector to detect the center of the object and enhance the feature description of the target object, based on a simple and efficient way for more accurate feature comparison. In particular, we introduce a motion-centric paradigm that localizes objects via motion in successive frame transformations. Experimental results demonstrate that our proposed method achieves satisfactory results on both the KITTI and nuScenes benchmarks, achieving a ~ 10% improvement in accuracy compared to state-of-the-art methods. Furthermore, our analysis confirms the effectiveness of each component and shows the great potential of the motion-centric paradigm when combined with representation matching.