View-invariant action recognition in surveillance videos

Abstract

Recently, human action recognition has been a popular and important topic in computer vision. However, except some conventional problems such as noise, low resolution etc., view-invariant recognition is one of the most challenging problems. In this paper, we focus on solve multi-view action recognition from surveillance video. To detect moving objects from complicated backgrounds, this paper employs improved Gaussian mixed model, which uses K-means clustering to initialize the model and it gets better motion detection results for surveillance videos. We demonstrate the silhouette representation “Envelope Shape” can solve the viewpoint problem in surveillance videos. The experiment results demonstrate that our human action recognition system is fast and efficient on CASIA activity analysis database.