3D Human Pose Estimation with a Catadioptric Sensor in Unconstrained Environments Using an Annealed Particle Filter.

Area of Science:

• Computer Vision
• Robotics
• Human-Computer Interaction

Background:

• 3D human tracking is crucial for human-computer interaction and robotics.
• Existing methods often rely on conventional cameras, limiting applicability in complex environments.
• Omnidirectional vision systems offer a wider field of view but present unique tracking challenges.

Purpose of the Study:

• To develop and evaluate a 3D human tracking system using catadioptric vision in complex environments.
• To address the limitations of traditional RGB camera-based tracking methods.
• To improve the robustness and accuracy of human pose estimation.

Main Methods:

• Utilized a particle filter framework combined with a catadioptric vision system.
• Employed Riemannian geometry for gradient computation on spherical images.
• Developed a robust descriptor using Support Vector Machine (SVM) classification for human detection.
• Proposed novel likelihood functions incorporating geodesic distances and silhouette overlap for particle filtering.

Main Results:

• Experimental evaluation on real-world data demonstrated favorable results.
• Achieved superior 3D pose accuracy compared to existing machine learning-based techniques.
• Measured a mean Root Mean Square Error (RMSE) of 0.065 m for 3D pose estimation during walking actions.

Conclusions:

• The proposed particle filter approach with catadioptric vision is effective for 3D human tracking in complex settings.
• The method offers improved 3D pose accuracy over conventional techniques.
• This research contributes to advancing human tracking capabilities in challenging, wide-view scenarios.