A method of simultaneously action recognition and video segmentation of video streams

Liang Ji, Rong Xiong, Yue Wang, Hongsheng Yu

Abstract: Human action recognition in video has been one of research hotspots in the field of computer vision due to its wide application. Normally video contains a variety of actions, most of the current action recognition methods are based on pre-segmented video clips, each containing only one kind of action. In this article, we propose a method of simultaneous segmentation and recognition for continuous videos. Based on the Bag of Words method, we extract features from data after the locality preserving projections in order to describe the different mode of actions. Then, to segment continuous video precisely, we make use of the multiclass SVM score and construct the problem as an optimization function solved via dynamic programming. Experimental results on Weizmann and UIUC datasets illustrate the benefits of our method compared to state-of-art approaches.