Incremental recognition of pedestrians from image sequences

The overall movement of articulated bodies (such as the human body) is enabled by the coordinated movement of its rigid body parts. The body parts are connected by joints and, in general, move di erently. For interpreting these movements as the movement of one single body, it seems to be necessary to incorporate knowledge in the analysis process. Therefore, in order to recognize pedestrians from monocular image sequences, we introduce a model-based approach. We represent the human body by a volume model and use medical motion data for simulating the movement of walking. This knowledge is exploited to determine the 3D-position as well as the posture of an observed person. By applying a Kalman lter we incrementally estimate the model parameters in consecutive images. Our approach has been tested on real image data.