Tracking-by-Counting: Using Network Flows on Crowd Density Maps for Tracking Multiple Targets.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Current multi-object tracking (MOT) methods predominantly use the tracking-by-detection paradigm.
• Detectors struggle with accuracy in crowded scenes due to occlusions and high object density.
• Existing methods are prone to errors in crowded environments or employ suboptimal two-step approaches.

Purpose of the Study:

• To propose a new multi-object tracking paradigm, tracking-by-counting, specifically designed for crowded scenes.
• To address the limitations of traditional tracking-by-detection methods in dense environments.

Main Methods:

• Developed a novel tracking-by-counting approach utilizing crowd density maps.
• Formulated a network flow program to jointly model detection, counting, and tracking.
• Achieved simultaneous global optimal detection and trajectory identification over entire video sequences.

Main Results:

• The proposed method demonstrates promising results across various public benchmarks.
• Successfully applied to diverse domains including people, cell, and fish tracking.
• Outperforms prior methods that do not adequately handle crowd density or use heuristic approaches.

Conclusions:

• The tracking-by-counting paradigm offers a more robust solution for multi-object tracking in crowded scenes.
• Jointly modeling detection, counting, and tracking via network flow optimizes performance.
• The approach shows broad applicability and effectiveness in challenging tracking scenarios.