Human activities discrimination with motion approximation in polynomial bases

In this paper we propose a new descriptor for human motion analysis. We try to recognize human motion represented by sequence of vector fields which are obtained with optical flow. Our method is based on an efficient measure of the similarity between two vector fields. This measure depends on the approximation of vector fields using projection onto a polynomial basis. From this similarity measure, we can retrieve a vector field within video sequences. As we consider a human motion as a sequence of vector fields, we use a simple algorithm in order to recognize motion. The results of the motion recognition system are shown for five boxing motions.