Recognition of Human Activity for Physical Rehabilitation from RGB-D Videos

In this paper, we present a 3D computer vision system for physical assessment and rehabilitation based on the data collected from a depth sensor. Physical rehabilitation is a complex and long-term process that requires clinician experts and appropriate tools. Hence a system that can successfully support the rehabilitation of individuals is of great interest to all the parties in sports and health industry. Exergames combined with activity monitoring technologies have been promising as a form of game that may make people involved in these video games more actively and physically. We propose a computer vision system for physical activity assessment. Due to the large temporal and spatial variations in actions performed by humans, human action recognition has been a long-standing challenge. Our approach recognizes certain human activities based on a motion descriptor that uses 3D human skeleton data. We define a motion descriptor (SHOJD) using the 3D distance between the most frequent key poses that occur throughout the action that is intended to be recognized. SHOJD features are then fed into an artificial neural network for classification. Experimental results indicate that the SHOJD based human action recognition system is robust with high recognition rate.