Experimental evaluation of Microsoft Kinect's accuracy and capture rate for stroke rehabilitation applications

To meet the challenges of ubiquitous computing for stroke rehabilitation, researchers have been trying to break away from traditional therapist-based modes of assessment. In this paper, the suitability of the Kinect to this end is experimentally evaluated. A set of thirteen gross movements, derived from common clinical stroke impairment level assessments (Wolf Motion Function Test, Action Research Arm Test, and Fugl-Meyer Assessment) were utilized to explore the Normalized Root Mean Squared Error (NRMSE) in position for data captured by Kinect as compared to a research-grade OptiTrack motion capture system. The specific joints of interest were the shoulder, elbow and wrist. A latency and capture rate estimation of the Kinect and its effects on data quality was also conducted. The NRMSE in position varied between 0.53cm to 1.74cm per data point among all axes and joints on average, when initial calibration was conducted via the OptiTrack system. The mean capture period was measured as 33.3ms with 3.86ms standard deviation, and the latency was observed to be on the order of two capture periods (66.6ms on average). Our results summarize the capabilities as well as limitations of Kinect in gross movement-based impairment assessment, in game-based rehabilitation paradigms, as well as in full-body motion capture applications in general.