Abnormal Events Detection Method Based On Antoencoder

With the rapid development of intelligent life,surveillance video equipment in public places have been continuously strengthened.Despite this,various emergencies are so numerous that security guards cannot detect anomalies and take effective measures in the first place.If we can effectively monitor the crowd and automatically detect abnormal events in the surveillance video,the relevant departments can respond and rescue promptly so as to minimize accidents and reduce loss of personal property.Therefore,video-based abnormal event detection has become a research hotspot in related fields.This paper detects abnormal behavior by improving the original autoencoder.This paper has conducted in-depth research on various abnormal behavior monitoring algorithms of domestic and foreign researchers.In view of the advantages and disadvantages of different methods,a detection method based on improved unsupervised learning is proposed.The improved automatic encoder is used to detect abnormal behavior and improve detection Recall.The main research contents are:1.An autoencoder model based on 3D convolution and 3D deconvolution is proposed.Traditional convolution layers use 2D convolution operations to extract image information,but for video,important time-domain feature information is ignored.In this paper,we propose to use3 D convolution and 3D deconvolution in autoencoders,and add the LSTM structure to the network.In the encoding part,3D convolution is used to extract temporal and space information,and in the decoding part,3D deconvolution is used to restore the original size and reconstructed image information.2.Skip connection.The skip-connection structure is one of the commonly used structures for image segmentation.Our algorithm draws on this structure to establish a connection between the encoding part and the decoding part of the autoencoder,and connects the feature map after the convolution of the encoding part with the feature map after the deconvolution of the decoding part.3.Object detection.Human detection is a key step in video content analysis.The method used in this paper is to extract HOG features and combine them with SVM classifiers for target detection.In order to improve the calculation speed,a KCF tracking algorithm is also used to detect targets at fixed intervals,and KCF is used as an aid in the middle frame.The test results show that the target can be effectively detected.4.Anomaly determination mechanism.After the reconstruction error is measured according to the above methods,the abnormal behavior score is calculated by using the reconstruction error to obtain a regularity score,and whether the abnormal state is determined according to the threshold.When abnormal behavior occurs,remind security personnel to intervene.In this paper,a number of experiments are performed on the datasets using this algorithm,and the final experimental results show that the methon is effective in monitoring abnormal behavior.Combining the two features of spatial and temporal domain to detect video target behavior,the accuracy and reliability are higher.When abnormal behavior occurs,a warning is issued to remind security personnel to be in place in time,thereby reducing the loss of personal and public property.