Discriminative and Efficient Label Propagation on Complementary Graphs for Multi-Object Tracking.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Object detection systems generate independent detections at each time instance.
• Associating these detections across time is crucial for multi-object tracking (MOT).
• Existing methods face challenges with sporadically available features and scalability.

Purpose of the Study:

• To develop a robust framework for associating object detections across time.
• To effectively utilize both spatio-temporal and appearance cues for improved tracking.
• To address the challenge of handling features that are only available sporadically.

Main Methods:

• Constructing graphs based on locally linear embedding of detection features.
• Defining graph neighborhoods to include temporally distant nodes with available appearance features.
• Formulating multi-object tracking as a difference of convex (DC) program.
• Decomposing the objective function into node-wise sub-problems for efficient computation.

Main Results:

• The framework successfully associates detections across time by propagating labels on constructed graphs.
• It uniquely leverages sporadically available appearance features for enhanced tracking accuracy.
• The proposed decomposition enables computationally efficient, scalable, and parallelizable solutions for practical tracking scenarios.

Conclusions:

• The developed graph-based approach provides an effective solution for multi-object tracking.
• The method demonstrates robustness in handling sparse appearance data.
• The computational efficiency and scalability make it suitable for real-world applications.