Model-free tracker for multiple objects using joint appearance and motion inference.

Area of Science:

• Computer Vision
• Artificial Intelligence

Background:

• Model-free tracking is challenging for multiple arbitrary objects due to detection and data association difficulties.
• Existing methods often fail with similar objects, appearance changes, and occlusions by updating appearance models independently.

Purpose of the Study:

• To develop a robust model-free multi-object tracking method that addresses limitations of existing approaches.
• To improve tracking accuracy and stability by jointly utilizing appearance and motion information.

Main Methods:

• Proposed a novel approach combining appearance and motion models learned jointly using graphical models and deep neural network features.
• Introduced an indicator variable to predict and manage sudden appearance changes and occlusions, preventing incorrect appearance model updates.
• Treated multi-object tracking as a graphical model inference problem, optimizing joint object locations simultaneously.

Main Results:

• The proposed method demonstrated superior performance compared to competitive multi-object tracking approaches.
• The joint learning of appearance and motion models, along with occlusion/change handling, led to more stable and accurate trajectories.

Conclusions:

• The developed method offers a significant advancement in model-free multi-object tracking, particularly in complex scenarios.
• Jointly learning appearance and motion models with robust occlusion handling provides a more effective solution for tracking arbitrary objects.