Tracking articulated objects in real-time range image sequences

A simple algorithm for tracking the pose of articulated objects in real-time range image sequences is proposed. This method models each target segment as a planar patch bounded by the convex hull of two circles, and utilizes both edge-like and region-like information in matching the model to the target. It uses hard constraints for joint attachment and is designed to be robust to occlusions and missing data. Experimental results are presented in which a human arm is successfully tracked over 26 frames of real, video-rate range imagery.