ORT: Occlusion-robust for multi-object tracking.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Multi-object tracking (MOT) struggles with long-term occlusions, where appearance features become unreliable.
• Existing methods like Kalman Filters fail in long-term motion prediction during occlusions.
• Data association in MOT is prone to identity switches due to similar scores during occlusions.

Purpose of the Study:

• To develop a robust MOT framework that effectively handles long-term object occlusion.
• To improve trajectory reconnection and reduce identity switches in occluded scenarios.
• To enhance motion prediction and data association for occluded objects.

Main Methods:

• Proposed a Long-term Spatio-Temporal Prediction (LSTP) module using spatial and temporal Transformers for state estimation and visibility prediction.
• Introduced an Ordered Association (OA) module that prioritizes trajectory association based on occlusion attributes and active states.
• Developed an Occlusion-Robust Tracker (ORT) framework integrating LSTP and OA modules.

Main Results:

• The ORT framework significantly alleviates the long-term occlusion problem in multi-object tracking.
• Achieved competitive performance on MOT17 and MOT20 benchmarks, outperforming state-of-the-art trackers.
• Demonstrated superior trajectory accuracy and purity, even without relying on appearance information.

Conclusions:

• The proposed LSTP and OA modules effectively address motion prediction and data association challenges posed by occlusion.
• The ORT framework offers a robust solution for multi-object tracking in scenarios with significant occlusions.
• This approach advances MOT capabilities by improving tracking robustness and reliability under challenging conditions.