A Kalman Filter based Approach for Markerless Pose Tracking and Assessment

The assessment and treatment of diseases that causes movement impairments typically rely upon clinical information obtained from self-reported rating scales and clinical observation from health professionals. Currently, objective clinical gait analysis requires the use of expensive cameras or wearable sensors, which may be too time-consuming for routine clinical usage. Therefore, an assessment system that can provide non-invasive gait analysis tool is desired. In this paper, we propose a cost-efficient assessment system that combines computer vision and artificial intelligence technology to analyse human gait, which can provide a basic clinical report for clinicians to evaluate the patients’ recovery to facilitate clinical decision making. And a series of experiments is taken to improve this assessment system performance. Those experiments showed that when the visibility threshold (VT) is set to a relatively high level (VT =0.4), the postprocessing part, which includes a Kalman filter and a FDF, can improve the human pose detection model (BlazePose)’s joint angle prediction accuracy by 10%. This post-processing method can be applied to other human body detection models to achieve filtering and feature extraction from joint angle signals for clinical gait analysis.