Human pose tracking in monocular sequence using multilevel structured models

Area of Science:

• Computer Vision
• Artificial Intelligence
• Human-Computer Interaction

Background:

• Accurate human body pose tracking in monocular video is crucial for numerous applications.
• Realistic scenes present challenges like background clutter, appearance variations, and self-occlusion.
• Tracking multiple individuals introduces further complexity due to inter-occlusion.

Purpose of the Study:

• To develop a robust method for hierarchical estimation of 3D human body poses from monocular video.
• To address challenges including automatic initialization, data association, and both self and inter-occlusion.
• To enable accurate multi-person pose tracking in complex, real-world scenarios.

Main Methods:

• A three-stage approach with multi-level state representation for hierarchical pose estimation.
• Stage 1: Coarse estimation of human positions and sizes via foreground blob tracking.
• Stage 2: 2D joint position inference using belief propagation on detected body parts.
• Stage 3: 3D pose inference using data-driven Markov chain Monte Carlo with 2D belief maps.

Main Results:

• The method successfully tracks multiple individuals in realistic indoor video sequences.
• Demonstrated capability in handling complex movements such as sitting and turning.
• Effectively manages self-occlusion and inter-occlusion between multiple people.

Conclusions:

• The proposed hierarchical approach provides a robust solution for multi-person 3D pose tracking.
• The method shows significant improvements in handling challenging real-world conditions.
• Enables reliable human pose estimation in applications requiring complex motion analysis.