Research And Application Of Video Anomaly Detection System Based On Memory Enhancement

With the continuous progress of modern technology and the development of optical imaging technology,it shows a rapid growth trend that video surveillance data address people’s needs for life and property safety effectively.At present,using deep learning methods for video anomaly detection has been believed to be a challenging issue in the field of computer vision.A research method,new and advanced,replaces many traditional methods based on extracting and analyzing video features of various dimensions.It is urgent that researchers hope to construct a model that can continuously learn high-dimensional features of normal behavior to achieve accurate and efficient detection of abnormal events,However,due to the extremely low frequency of abnormal events in daily life and they easily to being obstructed by other objects,this brings about a huge challenge for researchers.In most deep learning models used for video anomaly detection,the core module is the auto-encoder.However,owing to the strong representation ability of the deep auto-encoder after training in encoding and decoding,the output of abnormal samples after calculation by the model is not significantly different from the original image.Therefore,it is difficult to entrust all detection tasks to the auto-encoder.Based on this point,memory modules are utilized to memorize normal behavior patterns,enhance sensitivity to abnormal samples,reduce the impact of the strong representation ability of the auto-encoder,and improve the accuracy of detection.The research of this article is as follows:(1)A fully connected network based on memory enhancement is proposed.Inspired by the full connection layer of convolutional neural network,we propose a unsupervised learning method based on memory enhanced full connection network for video anomaly detection.In order to reduce the representation ability of the auto-encoder,we introduced an improved fully connected network based on memory modules.The training of the fully connected network module depends on the training results of the memory module,so we adopt a two-step training scheme to train the model.The experimental results show that this method has achieved good detection performance,with AUC values reaching 96.8%and 87.5%on two common datasets,UCSD Ped2 and CUHK Avenue,respectively,outperforming existing detection methods.(2)A bidirectional memory enhancement network is proposed.In order to accurately detect the occurrence of abnormal events and draw on the advantages of the automatic encoder model based on memory enhancement methods,we added skip connections between the feature maps generated by the encoder and decoder,and overlaid a memory enhancement module with bidirectional read and unidirectional update.Meanwhile,memory enhancement modules are added to each layer of the U-shaped structure and constructed a bidirectional memory enhancement network to decrease the "representation" ability of the automatic encoder.The improved model is compared on the public dataset and the impact of adding salt and pepper noise is analyzed on the experimental results,proving the effectiveness of the method.The AUC values on the two datasets reached 97.1%and 88.2%,respectively.(3)A video anomaly detection system based on memory enhancement is designed and implemented.The two improved methods mentioned above are used to train the network model for the deployment of the system.Firstly,there is a choice for users to use two different models based on their own situation,and then connect the monitoring video signal to the system.If an abnormality is detected,the system will issue an alarm message,providing reliable video anomaly detection services for users.