Exploiting Hierarchical Dense Structures on Hypergraphs for Multi-Object Tracking.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Current multi-object tracking (MOT) algorithms often fail with long-term occlusions and crowded scenes due to limited temporal windows and pairwise similarity considerations.
• Existing methods struggle to distinguish between spatially close targets with similar appearances.

Purpose of the Study:

• To develop a robust multi-object tracking algorithm capable of handling long-term occlusions and crowded scenarios.
• To improve the accuracy of distinguishing spatially close targets with similar appearances.

Main Methods:

• Formulating multi-object tracking as exploiting hierarchical dense structures on an undirected hypergraph based on tracklet affinity.
• Utilizing high-order similarities across spatio-temporal domains, considering appearance and motion globally.
• Employing a hierarchical optimization process for robust occlusion handling.

Main Results:

• The proposed algorithm demonstrates superior performance in distinguishing spatially close targets with similar appearances.
• The hierarchical design effectively handles long-term occlusions.
• Experiments on challenging multi-pedestrian and multi-face tracking datasets show favorable results compared to state-of-the-art methods.

Conclusions:

• The hypergraph-based approach offers a more effective solution for multi-object tracking, especially in complex scenarios.
• The algorithm provides a robust and accurate method for handling occlusions and crowded scenes in tracking-by-detection frameworks.