Local motion descriptors

Motion features play an important role in video retrieval. In this paper, we propose three new motion features that describe the local motion of arbitrarily shaped video objects. The proposed descriptors are based on the angular and circular area variances of the video object, the variances of the Angular Radial Transform coefficients of the object’s area, and the variances of the Fourier coefficients of the object’s boundary. These descriptors can be derived directly from the MPEG-4 compressed domain or can be computed using the binary shape masks of the video objects in the spatial domain. Our experiments show that the ranking obtained by querying with our proposed features closely match the human ranking.