Social Grouping for Multi-Target Tracking and Head Pose Estimation in Video

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Computer vision tasks often lack contextual information, leading to challenges in areas like multi-camera pedestrian tracking and head pose estimation from surveillance video.
• Visual ambiguities arise from varying pose, lighting, and low-resolution imagery, impacting the accuracy of these tasks.

Purpose of the Study:

• To integrate social grouping as novel contextual information into computer vision tasks.
• To enhance multi-target tracking and head pose/direction estimation in surveillance video using social context.

Main Methods:

• Developed a probabilistic formulation to model social grouping, multi-target tracking, and head pose estimation.
• Created effective solvers for the proposed probabilistic model.
• Utilized social grouping as a high-order association cue in single-camera multi-target tracking.

Main Results:

• Demonstrated that social grouping effectively mitigates visual ambiguities in multi-camera tracking and head pose estimation.
• Showcased social grouping as a natural high-order association cue for single-camera multi-target tracking, simplifying existing complex algorithms.
• Achieved performance improvements over models without social context and state-of-the-art approaches on public datasets.

Conclusions:

• Social grouping provides a valuable contextual cue that significantly enhances the performance of key computer vision tasks.
• The proposed probabilistic model and its solvers offer an effective solution for incorporating social context into visual analysis.
• The findings highlight the potential of leveraging social dynamics for more robust and accurate computer vision systems in surveillance and tracking.