3D single object tracking (SOT) based on point clouds is a fundamental task for environmental perception in autonomous driving and dynamic scene understanding in robotics. Recent technological advancements in this field have significantly bolstered the environmental interaction capabilities of intelligent systems. This field faces persistent challenges, including feature degradation induced by point cloud sparsity, representation drift caused by non-rigid deformation, and occlusion in complex scenarios. Traditional appearance matching methods, particularly those relying on Siamese networks, are severely constrained by point cloud characteristics, often failing under rapid motions or structural ambiguities among similar objects. In response, the research paradigm has progressively evolved toward motion-centric modeling approaches. These emerging frameworks utilize spatio-temporal joint modeling and geometric shape completion to attain notable performance gains. Furthermore, the incorporation of attention mechanisms and State Space Model (SSM) has enabled more effective multi-scale spatio-temporal feature association, which is particularly beneficial for long-term tracking scenarios. To the best of our knowledge, this is the first comprehensive survey dedicated to 3D single object tracking in point clouds. We provide a detailed analysis of current tracking methods, scrutinizing their limitations regarding multi-object interference and analyzing the trade-off between accuracy and computational efficiency. Finally, we discuss potential future directions, including the development of lightweight models for edge deployment and the integration of cross-modal fusion strategies.
Kuang et al. (Thu,) studied this question.
Synapse has enriched 5 closely related papers on similar clinical questions. Consider them for comparative context: