Sparse representation for action recognition

This paper presents an algorithm based on the ideas of bag of words and sparse representation for action recognition. We assume that all action instances form an action space and all action instances from one action class form a subspace of it. Furthermore, the action space can be represented by an over complete basis and each action instance can be represented by a linear combination of the basis. Naturally, the representation is sparse, so we can solve the problem via l¹-minimization. Then the action instance is recognized by how well the basis of one class represents it. Our algorithm is tested on the largest action dataset: KHT dataset. The result shows that our algorithm can work well within a relative small train set.