Research On Anomaly Detection Technology In Surveillance Video

With development of computer vision,How to processing surveillance video intelligent has become a pressing problem in security of Urban.Anomaly detection in surveillance video plays an increasingly important role as vital part of intelligent surveillance.The traditional anomaly detection model is divided into two types:abnormal segment detection and abnormal video frame detection.Although the traditional anomaly detection algorithm can detect abnormal conditions in the surveillance video,it still has the following problems: firstly,when processing the surveillance video segment,mostly methods use the three-dimensional convolution to extract the features,which contains whole information of video segment.But this method is very complex and needs lots of memory space and can't adapt to the changing scene,resulting in low accuracy of model detection.Secondly,in order to detect anomal in video frame,the current method use the encoder-decoder model to deal with this problem.To reconstruct the video frame requires a large amount of computing resources,which is also time-consuming.This thesis proposes two methods to detect anomalies in surveillance videos.The main work contents are as follows:Firstly,in order to solve the complexity of video pre-processing in Traditional model,we propose new method based on convolution long-term memory network FConv LSTM(Conv LSTM with Fusion Model).Instead of inputting entire video segment,this method samples video frames as input data.Then pre-trained model with model fusion is used to extract features from video frame.Due to multi video frame has time relationship,we use convolutional long-term memory network to correlate these video frame feature,so the space feature and time feature can be combine,which improve the generalization of model.Finally,the dataset is divided into sub-datasets to train model,combining these models can improve the performance of the model in abnormal detection.Secondly,in order to solve the low efficiency in video frame anomaly detection.In this thesis,Feature distribution based on product quantification(FDPQ)model is used to detect the abnormal video frame.in order to improve performance of model,we also use pre-trained model to extract feature,product quantification is used to Reduce dimensions,a normal video frame library is established to collect normal feature vector.When incoming video frame,after feature extraction and reduce dimensions,comparing video frame with the normal video library to get distribution distance.Normal scores is very low because of little change in video sequence,but abnormal video frames are obviously larger than others,so model can discriminate between abnormal videos frames and normal videos.The dimension reduction and feature distribution combine to process the video frame efficiently.