Multi-hypothesis motion planning for visual object tracking

Gong, Haifeng; Sim, Jack; Likhachev, Maxim; Shi, Jianbo · 2011 · OpenAlex-citations

DOI: 10.1109/iccv.2011.6126296

archive: archived pipeline: cataloged verified

Get this paper ↗ (DOI — opens at the source; we link to it, we don't host it)

Summary

This paper addresses the challenge of persistent occlusions in visual object tracking within crowded street scenes, where appearance changes and ambiguous data association often cause tracking algorithms to fail. The authors propose a long-term motion model based on multi-hypothesis motion planning, motivated by advances in robot motion planning. Unlike traditional models that rely on short-term predictions or single hypotheses, this approach constructs a set of "plausible" long-term plans for each person. These plans are goal-directed, obstacle-avoiding trajectories that avoid redundancies, unnecessary loops, and collisions. The core innovation is the efficient construction of disjoint plans across different homotopy classes, allowing the system to maintain multiple hypotheses for future paths during occlusions. The method utilizes winding numbers and winding angles to index homotopy classes of planned paths around obstacles. While previous work used complex L-values to distinguish homotopy classes, this approach replaces them with winding numbers (integers indicating loops around obstacles) and winding angles to create a more informative index. This allows the system to screen out implausible paths with excessive loops and handle obstacles of varying sizes effectively. The authors construct an augmented graph where vertices are equipped with winding angle vectors, enabling standard graph search algorithms like Dijkstra’s to find optimal paths within specific homotopy classes. In the tracking framework, the system maintains multiple hypotheses for each person. When visible, the person’s trajectory narrows the set of plausible goals; when occluded, the planner predicts re-appearance based on the plausible set of goal-directed paths. The algorithm is tested on a challenging real-world dataset collected from a car-mounted stereo camera driving through an urban city. The authors compare their method against Linear Trajectory Avoidance and a simplified linear planning model. The results demonstrate that the proposed multi-hypothesis motion planning algorithm outperforms both baseline algorithms in most sequences. By modeling pedestrian trajectories as goal-directed obstacle-avoiding paths with multiple hypotheses, the system successfully tracks individuals through complex occlusions and ambiguous appearances, such as when a pedestrian zig-zags to avoid cars and is subsequently occluded by another person with similar appearance. The significance of this work lies in its application of homotopy-constrained planning to visual tracking, providing a more expressive and realistic motion model than traditional dynamic social behavior models. By explicitly modeling multiple long-term hypotheses and avoiding redundant plans, the method improves robustness in crowded environments. The use of winding numbers and angles offers an efficient way to enumerate likely homotopy classes, addressing limitations in previous robotics planning methods that struggled with varying obstacle sizes and infinite search spaces. This approach enhances the ability to maintain consistent tracks over long occlusions, contributing to more reliable visual object tracking in complex real-world scenarios.

Provenance

The full processing record for this entry. Every stage of this paper's journey through the pipeline is logged — what ran, with which tool and model, how many attempts it took, and when it last completed.

Summary generated by qwen3.6-27b-prismaquant on 2026-06-26; verification: verified.

Topics

Ranked by relevance to this paper. Hover a topic for its definition.

• passenger motion sickness comfort