Accurate 3D Single Object Tracker in Point Clouds with Transformer

Abstract

3D single object tracking (SOT) in point clouds is a fundamental task in autonomous driving and robotics. Motivated by the success of transformer trackers in 2D tracking, we develop a 3D tracker in point clouds with transformer named Trans3DT. Trans3DT consists of three main designs: 1) Different from most previous trackers that use PointNet++ as the backbone, we propose a transformer feature extraction network to efficiently weigh point features to focus on deeper target cues. 2) Instead of using cosine similarity, we propose a novel target feature embedding network that fuses template and search area features in a global manner. 3) A voxel-to-BEV target location network is applied with a lightweight channel-related convolution block (CRCB) to enhance the BEV features. Extensive experiments demonstrate that Trans3DT achieves a new state-of-the-art performance on the challenging KITTI and nuScenes tracking benchmarks.