Target-Specific Action Classification for Automated Assessment of Human Motor Behavior from Video.

Area of Science:

• Computer Vision
• Human Motor Behavior Analysis
• Machine Learning

Background:

• Objective monitoring of human motor behavior is crucial for diagnosing and managing medical conditions.
• Wearable technology has advanced but struggles with complex, environment-interactive behaviors.
• Computer vision and deep learning offer new avenues for video-based behavioral analysis.

Purpose of the Study:

• To present a hierarchical vision-based behavior phenotyping method for classifying human actions from RGB camera videos.
• To address challenges in tracking multiple actors and classifying actions in dynamic, real-world environments.
• To enable objective assessment of motor behaviors through video analysis.

Main Methods:

• Implemented a cascaded pose tracker combining temporal detection relationships and appearance-based tracklet fusion for robust tracking.
• Utilized pose evolution maps derived from the tracker as interpretable, low-dimensional representations of movement sequences.
• Trained a convolutional neural network on pose evolution maps for action classification.

Main Results:

• The cascaded pose tracker achieved an average accuracy of 88% in tracking human actors.
• The overall system demonstrated an average test accuracy of 84% for target-specific action classification in untrimmed videos.
• The method effectively handles varying environments and fields of view.

Conclusions:

• The proposed vision-based method offers a viable alternative to wearables for complex motor behavior phenotyping.
• Accurate human action classification from video is achievable even in challenging, unconstrained conditions.
• This approach enhances the potential for objective, scalable assessment of human motor behavior.