VIDEO-BASED GAIT ANALYSIS BY SILHOUETTE CHAMFER DISTANCE AND KALMAN FILTER

Abstract

A markerless human gait analysis system using uncalibrated monocular video is developed. The background model is trained for extracting the subject silhouette, whether in static scene or dynamic scene, in each video frame. Generic 3D human model is manually fit to the subject silhouette in the first video frame. We propose the silhouette chamfer, which contains the chamfer distance of silhouette and region information, as one matching feature. This, combined dynamically with the model gradient, is used to search for the best fit between subject silhouette and 3D model. Finally, we use the discrete Kalman filter to predict and correct the pose of the walking subject in each video frame. We propose a quantitative measure that can be used to identify tracking faults automatically. Errors in the joint angle trajectories can then be corrected and the walking cycle is interpreted. Experiments have been carried out on video captured in static indoor as well as outdoor scenes.