Real-time estimation of human body postures using Kalman filter

Abstract

Presents a hybrid estimation method of human body postures from CCD camera images. In the hybrid estimation method, the feature points of the human body (top of the head, tips of the hands, and feet, and elbow joints) are obtained from the results of heuristic contour analyses of human silhouettes or those of a time subtraction image depending on the reliability of the silhouette information. A dynamic compensation is then carried out by tracking all feature points using the AR model in order to obtain their optimal position and to overcome self-occlusion problems. The AR model's parameters are estimated through online processing by the Kalman filter. The proposed method is implemented on a personal computer and the process runs in real-time. Experimental results show high estimation accuracy and the feasibility of the proposed method.