Sparse representation of adaptive key frame features for human action classification

Human action movement has constrained by the articulated body which leads to the variation of movement velocity from point-to-point. In this paper, adaptive key frame intervals are used to specify the proper number of frames by detecting the variation of human motion. Features which are extracted within the interval contain information of primitive movement which is similar among the same action. Then, the sparse representations of primitive movement are trained. The results on WEIZMANN demonstrate that the sparse representation within adaptive key frame interval can effectively classifies actions.