Integrating Scaling Strategy and Central Guided Voting for 3D Point Cloud Object Tracking

LiDAR-based 3D single object tracking has received remarkable attention due to its crucial role in robotics and autonomous driving. Most of them are based on hierarchical feature structures from PointNet++. However, existing based-stratified structure trackers ignore the fact that non-linearities in the narrow layers during backbone sampling corrupt the features. To resolve the problems, we propose an efficient 3D single object tracker with scaling strategy and center-guided vote enhancement (termed SCT), which can effectively maintain the integrity of point cloud features. SCT contains three novel designs: 1) A new feature extraction network is developed for 3D single object tracking, which proposes a 3D bottleneck separation module (BSM) that combines with the developed model scaling strategy to build hierarchical feature learning network. The BSM designs an inverted bottleneck design and separated multi-layer perception layers, which effectively reduces information loss or corruption. 2) A channel and spatial attention are then introduced into the feature fusion process to emphasize the potential key features in the fusion map. 3) A center-guided vote enhancement module based transformer is proposed to encode the position information of voting centers, and then adaptively assign weight to voting cluster features. Extensive experiments on KITTI and nuScenes benchmarks have shown that SCT achieves superior point cloud tracking in both performance and efficiency.