Visual Keywords’ Sparse Representation For Global Abnormal Events Detection

In many crowded public places, such as subway stations, railway stations and shopping malls, criminal behavior will cause great damage to the public security. Thus, public security attracts more and more attentions in the whole world. Video monitoring and abnormal events detection technology research become the main direction of this field. Based on computer vision, image processing, pattern recognition and other related technologies, We proposes two sparse representation methods for the global abnormal events detection. The main contributions of this thesis are as follows:(1)An improved histogram of maximal optical flow projection (HMOFP) method is proposed to extract features. We divide 0° -360° into 8 bins. The optical flow vectors with small amplitude are dropped to reduce time consumption. In the same bin, we project all optical flow vectors onto the angle bisector and select the maximal projection vector as this bin’s feature descriptor.(2)Over-complete dictionary construction is one of the key issues. Based on the feature similarity and the properties of histogram of maximal optical flow projection (HMOFP), we propose a new dictionary optimization algorithm. Firstly, we extract the HMOFP features of normal frames to construct the initial dictionary. Then, K-means algorithm is utilized to cluster the atoms into K classes. For all features in the same cluster, the maximum value of each row is selected to form a column vector. We treat this column vector as an atom of the optimized dictionary. Finally, K column vectors construct the over-complete dictionary.(3)Based on the feature similarity and images’ visual keywords, we propose the second dictionary construction algorithm. A frame is segmented into some patches and the HMOFP features are extracted from each patch. Then K-means algorithm is applied to cluster the features. The cluster centroids are cascaded to form a column vector and treated as an atom of the dictionary.(4)At last, we use OMP algorithm to obtain sparse representation coefficients and utilize sparse reconstruction cost (SRC) to detect abnormal events. If the SRC is more than or equal to the threshold, this frame is an abnormal frame. Otherwise, it is a normal frame.Experimental results show that for the global abnormal events detection, the detection accuracy of HMOFP feature sparse representation method is up to 91%. The accuracy of the visual keywords’sparse representation is up to 97%, which can effectively detect global abnormal events in crowded scenes.