Vision-based posture assessment to detect and categorize compensation during robotic rehabilitation therapy

Abstract

A vision-based posture assessment system for real-time monitoring of upper-limb robotic rehabilitation therapy is developed. The system is capable of automatically detecting and categorizing compensatory movements during robotic exercises and could be used in prompting the patient into the correct pose. A consumer depth camera and skeleton tracking algorithms were used to track the pose of the patient in real-time, and to extract a set of discriminating features which correlated with various posture modes. A multi-class classifier capable of incorporating temporal dynamics was trained to identify and categorize the most common types of compensation at high accuracy (86% per frame). A simple multi-stage active learning strategy was used to minimize the amount of manual annotation needed in providing the classifier with training data.