Research On Abnormal Event Detection In Intelligent Surveillance Video Based On Hybrid Features

Over recent years,massive adverse incidents often took place which caused great damages to the society.Social public safety could be effectively maintained through video surveillance.And it has become an integral part of our security system.There were disadvantages of low productivity and a lot of waste of resources in existing video surveillance system based on manual work,and it couldn't detect the potential danger in unattended situations and track the report,unable to play the active monitoring role in real time.These deficiencies motivated the continuous development of monitoring technology in the direction of intelligence.The detection of abnormal situations and alarms could be realized by intelligent video surveillance systems in a timely manner.So the potential safety hazards and loss of life and property of the people could be greatly reduced.Furthermore,the automatic analysis and detection of abnormal events is an important part of the intelligent video monitoring system.The current anomaly detection algorithm mostly used a single feature to detect that causes trouble about the scene feature description and had the shortcomings of the inaccurate description of the scene,which made the lower correct rate of detection of abnormal events.And there were some limitations in the method of detecting abnormal events by the change of the size of the velocity value.Therefore,this paper proposes an intelligent video anomaly detection algorithm based on hybrid features,which realizes automatic detection of abnormal events and has a higher accuracy and more practical.The hybrid features are obtained by combining the average velocity of the joint feature points,the average acceleration,the flat average energy and the texture features of the image.The main contents of the proposed method are as follows:(1)Detection of moving target.Firstly,in order to satisfy the constraints of the LK optical flow method,the Gaussian pyramid was used to reduce the size of the sequence image and the speed of the moving object.Then the LK optical flow method combined with the Gaussian distribution was used to segment and detect the moving object.(2)Extraction of target hybrid features.Firstly,Harris corner detection method was used to extract the spatial corner feature quantity;Secondly,the extracted corner feature was to calculate the optical flow characteristics of the sparse feature set combined with the pyramid LK optical flow method,and then the average of speed,acceleration and energy of feature points in each frame were obtained through further solving.The average of speed,acceleration and energy were taken as the 3-dimensional motion feature quantity;Finally,a gray-level co-occurrence matrix was created to extract the texture feature quantity.A Gaussian model algorithm and a fast Fourier transform were used to improve the accuracy and the extraction efficiency of texture features.Then the leave-one-out method and SVM classifier are applied to optimize texture features.Combination of the energy,the entropy,the moment of inertia and the correlation makes contribute the most to the accuracy of classification.And their mean and standard deviation were taken as the 8-dimensional texture feature quantity to represent edge texture information for different video frames.Finally,the 3-dimensional motion features quantity and the 8-dimensional texture feature quantity were combined and used as the final 11-dimensional hybrid features quantity for abnormal event detection.(3)The detection of abnormal event.The automatic detection of abnormal events was implemented by training the representative video images which used a surveillance video anomaly detection algorithm based on hybrid features and an SVM classifier,and the correctness of the 11 hybrid features quantities selected in the previous text and the feasibility of the algorithm were verified.The result demonstrates that the anomaly detection algorithm proposed in this paper can effectively detect anomalous events with a detection rate of 97.65%.And compared with the detection of abnormal events based on a single feature,the new algorithm can detect anomalies and achieve better performance.