Description and tracking of moving articulated objects

This paper discusses the problem of deriving a description for an articulated object from images taken in the real environment. Sometimes with the articulated object, the detection of the articulation is difficult depending on the deformation or the direction of the view line. Also, the false articulation is detected due to the disconnection of the object contour derived from the image or due to the noise. To cope with such difficulties, a method is proposed for the moving articulated object, where the articulation is detected with a high reliability, and the description as well as the segmentation of the articulated object are derived by processing a series of images containing different outblocks and deformations of the object. In this method, the ribbon, which is the two-dimensional version of the generalized cylinder, is used as the basic representation for the part. The initial description is derived from each frame image, and the initial descriptions of various frames are compared by ribbon matching to detect the articulation position with a high reliability. Based on the detected positions, the initial descriptions are integrated selectively and the final description is obtained. As a by-product of the ribbon matching, the tracking of parts also is realized. By an experiment using a human walking scene, the usefulness of the proposed method is demonstrated.