Abstract
3D Single object tracking (SOT) in point clouds is a computer vision task that involves localizing the target object in subsequent frames given the initial state. Most 3D SOT methods have only employed the current feature fusion map while overlooking the significance of temporal knowledge. However, it is crucial to incorporate temporal knowledge of the previous target cues into the current feature fusion map, especially in situations involving occlusion, significant appearance variations, and sparse point cloud representations of the target object. In addition, the performance of object tracking is highly dependent on a discriminative feature fusion map. To address these issues, we propose a 3D SOT by exploiting temporal knowledge and attention mechanisms. First, we design a Cascading Attentionbased Feature Fusion (CAFF) module to integrate the target cues into the search area, taking into account the intra-feature dependency within each feature and the inter-feature dependency between the template and search area features. Second, we present a Temporal Relation Modeling (TRM) module to refine the current feature fusion map by propagating temporal target cues from the previous feature fusion map. To further enhance the performance of the tracker, we present a template update sub-module to address the challenges posed by low-quality templates and appearance variations. The experimental results obtained from the KITTI, nuScenes, and Waymo Open Dataset (WOD) datasets demonstrate that the proposed approach achieves competitive performance compared to state-of-the-art methods. Our method runs at 67.57 FPS on a single RTX 2080 GPU.
Funding Information
  • National Natural Science Foundation of China (62471107, 62461053)