Recognizing human actions by fusing spatial-temporal HOG and key point histogram

This paper presents a novel human action recognition method which fuses spatial-temporal histogram of oriented gradient (STHOG) and key-point histogram (KPH) features. First, we employ a cuboid detector to extract the interest points from a video clip. Then for each interest point, a cuboid around the point is extracted and the STHOG of the cuboid is calculated. In addition, all the three-dimensional (x, y, t) interest points of a video clip are quantized to form a histogram of key-point distribution. By concatenating the STHOG with the KPH, we obtain a novel spatial-temporal descriptor, which is used to classify the action types. Simulation results indicate the proposed method improves recognition power.