Multiobject Tracking by Submodular Optimization.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Optimization

Background:

• Multiobject visual tracking is crucial for various applications.
• Occlusion poses a significant challenge in tracking multiple objects.
• Existing methods often struggle with robustly handling occluded targets.

Purpose of the Study:

• To propose a new multiobject visual tracking algorithm.
• To address the challenge of object occlusion in tracking.
• To leverage submodular optimization for improved tracking accuracy.

Main Methods:

• A two-stage approach involving tracklet selection and trajectory formation.
• Generating candidate tracklets using overlap criteria and min-cost flow.
• Formulating the tracking problem as a submodular maximization problem with constraints.
• Designing a connecting process to resolve occlusions.

Main Results:

• The proposed algorithm effectively handles occlusions.
• Experimental results validate the algorithm's superior tracking performance.
• The method demonstrates robustness in complex tracking scenarios.

Conclusions:

• The submodular optimization-based tracking algorithm offers an effective solution for multiobject visual tracking.
• The proposed tracklet selection and trajectory connection strategies significantly improve occlusion handling.
• The algorithm provides a promising direction for future research in visual tracking.