Indoor And Abnormal Behavior Detection Algorithm Based On Svm

Nowadays, most of video surveillance systems focus on detecting or tracking the moving targets in the scenes, However, the purpose of monitoring is to detect and analysis abnormal events or abnormal behavior monitored in the scenes. Manual handling is neither practical nor economical for video sequences in long time. Therefore, abnormal detection carried out in the video surveillance is very important and necessary.This thesis studies abnormal behavior detection in indoor environment. Based on the silhouette and the front feature of human, respectively, abnormal detection algorithms are investigated. In this thesis, the support vector machine (SVM) classifier is introduced, and the training and the test of the classifier is carried out on a home-brewed dataset.For the anomaly detection from the silhouette of human, in the thesis, falling down is regarded as an abnormal behavior that is subjected to detection. Some characters are used, which are the Centroid track and the contour width of human. The SVM classifier with linear kernel is used to identify the abnormal behavior. Experimental results show that the correct recognition rate is more than 97% and the false alarm rate is low, meanwhile, the computational cost is relatively low.Based on the front abnormal detection, a new feature extraction method based on grid representation technology is used. First, the body contour is mapped to a two-dimensional array; then, sampling feature points on the silhouettes, the relative location information of these points constitute a feature vector; finally, we use an SVM classifier for identifying abnormal behavior, and an encouraging performance has been achieved.All above mentioned experimental results are achieved through an abnormal behavior detection system that developed on Visual C++6.0 and DirectShow by ourselves.