Research On Abnormal Behavior Detection On Video

Abnormal behavior detection in surveillance video is an important research direction in the field of intelligent video surveillance.Abnormal crowd behavior detection in surveillance which has wide application prospects in real life is able to predict or deal with emergencies promptly,while saving the cost of manual detection.Therefore this issue is concerned by a growing number of researchers.It is of great theoretical and practical significance to carry out a thorough study of it.Firstly,abnormal behaviors of pedestrians or substance are described as abnormal objects in this paper.In order to provide the basis of abnormal objects basic presentation,extracting features and classifier design,Theoretical knowledge correlated with abnormal objects detection is studied including background modeling,low-level image features and convolution neural network.Then research work in this paper is divided into two parts as follows.(1)To solve the problems that panic behavior is difficult to obtain and the apparent characteristics are not obvious,a crowd panic behavior detection algorithm based on motion effect map features of moving foregrounds is proposed.Firstly,moving foreground segmentation algorithm based on adaptive GMM model is used to extract the foregrounds of the video sequence.Then each video frame is divided into blocks in order to achieve motion effect map features of moving foreground blocks by acquired foreground area.An improved K-means algorithm by optimizing initial clustering centers is employed to train and test dataset.The experimental result shows that the proposed method maintains quite great performance when detecting panic behavior in outdoor scene,while the AUC of the indoor panic detection increases by at least 2.64%,reaching 86.85%,which effectively improves the accuracy of detecting unusual panic behavior compared to existing algorithms.(2)To detect the motor vehicle,rider and skater intruded on the pavement,an anomaly detection algorithm based on joint double stream convolution neural network is proposed in this paper.Firstly,moving foreground segmentation algorithm based on adaptive GMM model is used to extract the foregrounds of the video sequence.Lots of detecting windows are obtained through Non-maximum value suppression from fixed-size candidate windows which took foreground pixels as center.Secondly,a convolution neural network(CNN)described as improved CIFAR-10-Full model is used to model the single image from detecting window.Similarly,another improved CIFAR-10-Full is used to model the optical flow amplitude diagram from detecting window.In the finally detection stage,detecting results of two independent CNN model are combined by the strategy of weighted sum rules.The experimental result shows that the AUC of proposed method increases by at least 7.59% and reaches 98.46% compared with the existing algorithm,while the location of the exception object can be located.