Research Of Crowd Abnormal Event Detection Based On Fully Convolutional Neural Network

In recent years,the occurrence of group violence has caused more and more attention to the safety problem of public places,but effective monitoring of public safety has been a difficult problem.Analyzing monitor videos in an artificial way often ignores a certain detail and cannot locate the time and area of the accident accurately.It not only wastes manpower and material resources,but also indirectly increases the harm caused by danger.Therefore,the urgent need of high accuracy video surveillance system makes crowd abnormal event detection be widely concerned.Abnormal event detection in group scenario is a challenging project in the field of computer vision.The difficulty is that the definition of anomaly is not strict,and abnormities vary widely with scene changes.Many scholars have put forward many models and algorithms,which have made good progress.However,most of these algorithms are traditional methods based on hand-craft features,which not only require complex preprocessing,but also perform poorly on the efficiency of the algorithm.Based on the background,this paper will use a new method based on deep learning to detect the abnormal events in the videos.The main work is as follows:(1)An algorithm of anomaly behavior video feature extraction based on Fully Convolutional Neural Network(FCN)is proposed.Due to the limitations of traditional methods in crowd abnormal event detection,the pre-training CNN model is converted into FCN,which is used to efficiently extract the features of video frames.Then,in order to further improve the efficiency of the algorithm,the Iterative Quantization(ITQ)method is added to quantize the convolutional features.And the ITQ method is used as the feature encoding layer,which is combined with FCN to form a new feature extraction process of videos.(2)A video anomaly measurement algorithm based on the histogram of binary feature(HBF)is proposed.Many crowd abnormal detection algorithms identify abnormal events by training a classification model,which not only requires a large amount of prior knowledge,but also increases the complexity of the algorithm.Aiming at the problem,this paper defines a quantitative attribute: abnormal coefficient,to analyze the abnormal degree of video intuitively.HBF is based on the convolution feature of video images,which does not depend on the precise calculation of the trajectory,and is more robust and more concise.(3)After using HBF to detect the global anomaly,a method integrating abnormal coefficient and optical flow information is proposed to locate the abnormal area.Finally,tests are carried out on standard databases for group anomaly detection to verify the effectiveness and real-time performance of the proposed method.The experimental results show that compared with other traditional methods,the FCN-HBF method can detect the abnormity in video more quickly while ensuring a certain accuracy.