Group Anomaly Event Detection Based On Surveillance Video

With the rapid development of Internet technology,the traditional video surveillance technology has been constantly improved.Intelligent video surveillance has become an inevitable trend and has got people's attention and recognition.In recent years,intelligent video surveillance has gradually penetrated into all areas of society.Some key technologies among them such as face recognition,gait recognition,and anomaly detection have been widely applied in airport,train station,subway and other important places,which have become an important means to protect social public safety.Group abnormal event detection as the primary task of intelligent video surveillance,its main job is to monitor the abnormal events in the video timely and accurate early warning and detection,to bring the greatest possible reduction in the harm caused by accidents.The researches of group abnormal event detection are mainly divided into two aspects:basic event representation and detection model establishment.This thesis is from those two aspects to study the abnormal event detection.The main work is as follows:(1)In terms of basic event representation,this thesis mainly studies the representation of low-level visual features based on events and put forward the significant new optical flow histogram descriptor.Specific operations are:firstly to extract the optical flow information of the video image blocks(i.e.optical flow vector)and divides it into different directions.Thus,the direction information of the optical flow is classified;Secondly,Thresholds are set for each direction interval and discard the optical flow with the amplitude less than the threshold value;only the remaining large-scale optical flow information is counted.Calculating the sum of the optical flow values in each interval,thus we get the intensity information in different directions within the range of the optical flow histogram features;finally,the statistical histogram constitutes the feature as the feature descriptor of the image block.The feature descriptor is able to classify the optical flow information more flexibly by setting the threshold of the optical flow scale.It can express the motion information in video very well.(2)In the aspect of detection model building,this thesis proposes an improved anomaly detection model based on sparse representation.It is called anomaly detection model based on clustering.Specific operations are:firstly,an initial dictionary and the corresponding coefficient matrix are obtained from the training samples;Then clustering the initial dictionary based on the coefficient matrix to form a plurality of sub-dictionaries,and optimizing the respective sub-dictionaries to obtain a multi-dictionary combination.Each of these sub-dictionaries represents a class of normal events.Finally,according to the reconstruction error of the test sample in the multi dictionary combination,we can judge whether it is an abnormal sample.The main improvement of the model is the clustering dictionary learning algorithm.It translates a traditional over-complete dictionary into multiple sub-dictionaries.So in the detection stage,for a given sample,we just need to find the most suitable sub-dictionary and calculate the test sample in the sub-dictionary under the reconstruction error to determine whether the test sample anomalies.(3)In this thesis,the proposed algorithm is verified and analyzed by experiments on the two public datasets.Experiments show that the method proposed in this thesis has better detection performance compared with other methods when it is used to detect the abnormal events of groups in video surveillance under the crowded scene.