On-Line Visual Tracking with Occlusion Handling

Area of Science:

• Computer Vision
• Artificial Intelligence
• Robotics

Background:

• Occlusion presents a significant challenge in online visual multi-target tracking, limiting algorithm performance in real-time applications like surveillance.
• Existing online tracking methods struggle with occluded targets due to their reliance solely on past and present data.

Purpose of the Study:

• To develop an advanced online algorithm for visual multi-target tracking that robustly handles occlusions.
• To improve the accuracy and continuity of target tracks in challenging visual environments.

Main Methods:

• Extension of the Generalized Labeled Multi-Bernoulli (GLMB) filter for visual multi-target tracking.
• Implementation of a novel information propagation mechanism to create a 'déjà vu' effect using visual and motion features.
• Inclusion of advanced false alarm detection/removal and label recovery techniques.

Main Results:

• The proposed GLMB filter extension demonstrates superior performance in handling occlusions compared to state-of-the-art methods.
• Significant improvements observed in metrics such as reduced ID switches and track fragmentation.
• Effective recovery of tracks even after substantial periods of occlusion.

Conclusions:

• The developed GLMB-based algorithm offers a robust solution for online visual multi-target tracking under occlusion.
• The novel approach enhances tracking reliability and accuracy, crucial for applications in video surveillance and sports analytics.