Action and gait recognition from recovered 3-D human joints

Area of Science:

• Computer Vision
• Biometrics
• Machine Learning

Background:

• Human action and gait recognition are crucial for surveillance and human-computer interaction.
• Existing methods often require manual feature engineering or specific viewpoints, limiting their applicability.

Purpose of the Study:

• To develop a unified, viewpoint-free framework for simultaneous action recognition and person identification.
• To leverage markerless 3D pose recovery for robust feature extraction.

Main Methods:

• Markerless 3D pose recovery to extract joint and pose parameter sequences from volume data.
• Extraction of configuration (joint combinations) and movement (body dynamics) features.
• Modeling features using Hidden Markov Models (HMM) and exemplar-based HMM.
• Hierarchical classification fusing movement and configuration features for action recognition and gait-based person identification.

Main Results:

• Demonstrated effectiveness on the Institut National de Recherche en Informatique et Automatique Xmas Motion Acquisition Sequences dataset.
• Achieved accurate action classification and person recognition through fused feature modeling.

Conclusions:

• The proposed viewpoint-free framework offers a robust and integrated approach to action and gait recognition.
• Markerless 3D pose recovery combined with HMM-based feature modeling provides an effective solution for human behavior analysis.