A new simple human abnormal action detection based on static images

Human abnormal action is an active issue in the computer vision domain. Most of the current approaches rely on spatio-temporal feature. A great deal of work has been done to show the feature with good performance, but expensive computation. There are some applications with low requirements for human abnormal action detection. We propose a new simple but low computational human abnormal action detection method for that applications. First, static images are extracted with equal interval along time. Then, images are divided into patches and use distance transformation to get feature vectors. Normal action patches are classed and represented by cluster centers. Experiments are conducted on the well known KTH dataset and a video dataset we recorded to show the efficacy of the proposed method.