Action Recognition Based on Sine Series Fitting

This paper presents a sine series-based method for action recognition. The proposed method can be divided into three stages: feature extraction, series fitting and feature matching. In the stage of feature extraction, the method gets binary human silhouette through background difference and Gaussian background modeling at first, then uses the binary silhouette to represent the given image sequence and calculates the distance from the silhouette centroid to the boundary points according to a clockwise movement, thus changing the human silhouette into distance curve. In the stage of series fitting, the method fits the curve with the sine series and changes the distance curve into sine parameters with different amplitude, frequency and offset, which reduces the computational cost greatly and changes the process of action recognition into the matching of curve parameter features. Finally, in the stage of feature matching, the method classifies each frame in the given image sequence through calculating the minimum distance between it and the known action categories at first, and then gets the category of the given image sequence by voting. We adopt a leave-one-out scheme for experiment evaluation. The experiment results on a database of 90 short video sequences show that the promising performance is both effective and efficient.