Point-Voxel Guidance Fusion With Unidirectional Interaction for 3D Single Object Tracking

Abstract

Sparse point-based trackers struggle with textureless and incomplete point clouds. Conversely, dense voxel-based trackers have richer spatial and semantic information, but how to filter out interference from complex backgrounds remains a challenge. Besides, there is still a gap between point and voxel based trackers to exploit their complementary strengths. To address these issues, we propose a novel point-voxel guidance fusion tracker with unidirectional interaction, denoted as PVTrack, which constructs the bridge between point and voxel tracking features to complement and enhance each other. Specifically, we design template-enhanced unidirectional attention (TEUA) and historical template fusion (HTF), which enable unidirectional interaction from historical templates to the search area in point branch, retaining the pure template features. Then, a point-guided adaptive feature transformer (PGAFT) is developed to dynamically enhance the interaction between point and voxel features from a global perspective. Finally, we design a motion feature module (MFM) to capture the historical motion data, which not only refines target movement but also provides a motion prior. Extensive experiments demonstrate that PVTrack achieves superior performance, reaching an average accuracy of 89.5%, 72.58%, and 63.4% on the KITTI, NuScenes, and Waymo Open Dataset, respectively.