Abstract: Aiming at the problem of the global action feature that is difficult to accurately extract, this paper uses space-time interest points to represent motion. In view of the great error in the traditional quantizative word bag hard classification method, this paper refers to the idea of fuzzy clustering and proposes a soft classification approach. Firstly, interest points detect algorithm is applied to extract visual words from the video, and this paper builds the codebook via K-means clustering. This paper calculates the distance of the visual words to be classified and each codeword in the codebook and then gets the membership probability to each codeword. Finally, the codeword’s frequency in each video clip can be calculated. The performance is investigated in Weizmann and KTH datasets. The experiment result shows that the average recognition rates increase 8% in Weizmann datasets and 9% in KTH datasets. This proves that the approach can recognize human behavior more effectively.

Key words:  algorithm, interest point, action recognition, fuzzy clustering