A novel hierarchical data association with dynamic viewpoint model for multiple targets tracking

• A new online multiple targets tracking framework is proposed.
• An iterative algorithm on DVM filters out some false alarms.
• Incremental motion pairing inference strategies exploit motion similarities.

A new framework of hierarchical data association tracking (HDAT) with branch partition, candidate upgrading and incremental motion pairing inference is proposed to resolve the problem of online multiple targets tracking. Branch partition divides the process into several independent parts so as to reduce the computational complexity on affinity. Candidate upgrading improves the robustness of target initialization by tracking potential targets and incremental motion pairing inference could benefit the occlusion handling. Furthermore, a dynamic viewpoint model (DVM) and its iterative computation algorithm are developed for tracking multiple targets under moving camera videos. Extensive data experiments on several public benchmarks show that the presented approach achieves comparable results to state-of-the-art on static camera videos and promising results on moving camera videos, and moreover, the runtime performance is significantly improved.

Multiple targets tracking by hierarchical data association combined with dynamic viewpoint model (detection, targets, and candidates are marked by blue rectangles, red ellipses, and green hexagons, respectively).Figure optionsDownload high-quality image (78 K)Download as PowerPoint slide