A poselet based key frame searching approach in sports training videos

In some sport training application, it is necessary to search the key frames of training video for carefully analysis. In this paper, we take the key frame searching issue as a pose estimation problem. First, a set of various pose detectors are collected trough the twice SVM training process, each of which can be interpreted as a learned pose-specific HOG weight classifier. Then we run each linear SVM classifier over the image in a multi_scale scanning mode. In order to resolve the problem of extreme similarity between the adjacent frames, the detection hits at every scale in each frame is counted as the principle of optimal key frame selection. The frame with the most detection hits are chosen as the key frame for the pose detector. The experimental results using weight-lifting training videos show the efficiency of proposed approach.