Inverse Dynamics for Action Recognition.

Area of Science:

• Computer Vision
• Robotics
• Biomechanics

Background:

• Pose-based human action recognition traditionally relies on kinematic features.
• Kinematic features present challenges like high dimensionality and limited inter-class variation.
• Existing methods often neglect environmental interactions influencing human motion.

Purpose of the Study:

• To propose a novel method for human action recognition using dynamic features.
• To leverage inverse dynamics on a physics-based human body model.
• To achieve more discriminative and lower-dimensional action representations.

Main Methods:

• Developed a physics-based, articulated human body model with muscles and variable joint stiffness.
• Derived dynamic features, including joint torques (hip, knee), gravity, and ground reaction forces.
• Utilized these dynamic features within a Hidden Markov Model (HMM) for classification.

Main Results:

• Dynamic features provide a more discriminative and lower-dimensional representation compared to kinematic features.
• The proposed method achieves good classification performance with limited training data.
• Effectiveness validated on CMU motion capture and Osaka University Kinect datasets.

Conclusions:

• Dynamic features derived from inverse dynamics offer a superior approach to human action recognition.
• The method's efficiency and accuracy are demonstrated across diverse datasets and actions.
• This approach enhances motion understanding by incorporating physics-based dynamics.