Boost Action Recognition through Computed Volume

We detect interest points in temporal-spacial space and use the local feature plus their positions to recognize action in a video. Although some previous methods take advantage of the position of each interest point besides the local feature of them, and achieve good performance, it consumes much time to position these points due to the complexity of an action. We propose two simple methods to position each interest point, and design a new feature for action recognition. Evaluation of the approach on two sets of videos suggests its effectiveness.