Crowd Abnormal Event Detection In Public Monitoring Scene

With the popularization of many projects such as "Safe City" and "Safe Campus" all over the world,intelligent video surveillance plays a crucial role in the field of public security.And Abnormal event detection is the core of improving the level of intellectualized degree of the video surveillance system,which make great influence on public safety and other fields.The occurrence of abnormal events with a small probability in surveillance video,which is diverse and difficult to obtain.It is difficult to obtain equivalent normal events and abnormal events,At the same time,Different abnormal events with different occurrence probability.The imbalance of data is the difficulty in training a good classification model.consequently,the main contents of this paper are as follows:1)Abnormal events with a small probability.These low-probability events include abnormal appearance,abnormal energy,and abnormal directions.The normalized pixel information is taken as the appearance feature of the object and The membership value corresponding to the optical flow energy is used as the energy characteristic of the object.The membership value corresponding to the optical flow direction is used as the movement characteristic of the object.After they are linearly fused,the information of the object can be better expressed.2)Aiming at the phenomenon of missing detection in the low-probability anomaly events and The optical flow information contains more abundant motion information of the object,However,anomalous events are indefinite event and cannot be accurately determined by optical flow information.Therefore,the concept of membership of optical flow energy and the degree of membership of the optical flow direction are introduced in this paper.The relationship between the optical flow distribution and the corresponding frequency in the training data is statistically calculated.It is highly likely that the optical frequency with extremely high frequency is the normal event.The membership function can be fitted by corresponding degree of membership value.From this function,the membership value of any optical flow value and direction belonging to the normal event can be obtained.3)Aiming at the problem of unbalanced data in abnormal event detection,this paper use convolutional autoencoder network to model the normal events which is easily extracted.The input of the network is not only a single frame feature,but also a combination of consecutive image frame features as input.After the temporal convolution operation,the motion information can be effectively encoded into the network structure.The trained model can well reconstruct normal events,In the test phase,the trained model is used to reconstruct the test data.The smaller the error is,the more likely it is a normal event.Otherwise it is abnormal event.This method can effectively solve the problem of unbalanced data.The experimental results of the algorithm in the standard data set verify the effectiveness of the proposed method in detecting abnormal events with low probability.