Abstract: The acquisition of specific motion frames in motion video was an important part of intelligent teaching. In order to obtain specific motion frames in video for further analysis, a method of extracting specific motion frames from motion video was proposed using the knowledge of pose estimation and clustering. Firstly, the HRNet attitude estimation model was adopted as the basis, which was of high precision but large scale. To meet the needs of practical application, this paper proposed a Small-HRNet network model by combining it with the data encoding of DARK. The parameters were reduced by 82.0% while the precision was kept unchanged. Then, the Small-HRNet model was employed to extract human joint points from the video. The human skeleton feature in each video frame served as the sample point of clustering, and finally the whole video was clustered by the skeleton feature of the standard motion frame as the clustering center to produce the specific motion frame of the video. The experiment was carried out on the martial arts data set, and the accuracy rate of the martial arts action frame extraction was 87.5%, which can effectively extract the martial arts action frame. 

Key words: specific motion frame, attitude estimation, data encoding and decoding, movement characteristics, clustering