Multiple target tracking by learning-based hierarchical association of detection responses

Area of Science:

• Computer Vision
• Artificial Intelligence
• Robotics

Background:

• Multiple target tracking from a single camera presents challenges in maintaining target identities and handling fragmented trajectories.
• Existing methods often rely on heuristic model selection and manual parameter tuning, limiting adaptability and performance.

Purpose of the Study:

• To develop a robust hierarchical association approach for multi-target tracking from a single camera.
• To improve tracking accuracy by minimizing identity switches and reducing trajectory fragmentation.
• To introduce an automated method for learning tracklet affinity models without manual parameter tuning.

Main Methods:

• A conservative dual-threshold method generates initial tracklets with minimal identity switches.
• Tracklet association is formulated as a Maximum A Posteriori (MAP) problem, solved using the Hungarian algorithm.
• A novel bag ranking method and boosting algorithm are used to learn tracklet affinity models from various features.

Main Results:

• The proposed hierarchical association approach significantly enhances tracking accuracy in challenging multi-pedestrian tracking scenarios.
• The method demonstrates a considerable reduction in tracklet fragmentation and identity switches compared to state-of-the-art algorithms.
• Learned tracklet affinity models provide more accurate associations, improving overall tracking performance.

Conclusions:

• The hierarchical association framework offers a superior solution for single-camera multi-target tracking.
• Automated learning of tracklet affinity models through the novel bag ranking method improves robustness and accuracy.
• The approach effectively addresses key challenges in multi-object tracking, particularly for pedestrian datasets.