Video Anomaly Detection Based On Predictively Reconstructed Frames

In recent years,intelligent surveillance video anomaly detection has been widely used.However,the current anomaly detection methods are still insufficient.Anomaly detection based on supervised has huge manual workload.Anomaly detection based on image reconstruction often fails to specifically confirm the classification of normal anomalies,resulting in poor classification effect;Anomaly detection based on prediction has poor image quality and large amount of calculation.The research on video anomaly detection algorithm carried out in the academic field has poor effect when applied to the specific data set of practical application.The research content of this thesis is divided into three parts.The first one is video anomaly detection algorithm of continuous frame prediction and image reconstruction based on residual attention mechanism.the second one is the generative adversarial network anomaly detection algorithm based on double discriminator and the classification threshold function algorithm based on two-dimensional space.Last,we design and implement the video anomaly detection software based on frame image reconstruction prediction.The main work of this thesis is as follows:1.Aiming at the problem of prediction and reconstruction of video continuous frame images,a video anomaly detection algorithm based on prediction frame is proposed.We extract image features and generate predicted images by the idea of generative adversarial network,u-net network,residual idea and attention mechanism.We integrate optical flow loss,intensity loss,gradient loss,mean square error loss,and structural similarity loss functions to perform multi-scale video anomaly analysis on surveillance video.Compared with similar methods,the AUC of this anomaly detection method is improved by 1.2%～1.9%on different data sets.2.For the problem that the accuracy of anomaly discrimination is not high enough,this thesis proposes an anomaly detection algorithm of generative adversarial network based on double discriminators and a classification threshold function algorithm based on two-dimensional space.The double discriminator anomaly detection algorithm discriminates the local features and the overall features at the same time,so as to improve the quality of the image generated by generative adversarial network.The classification threshold function algorithm takes the classification results of two different network models as the two eigenvalues of the image,and uses support vector machine for adaptive scatter classification,so as to obtain the classification threshold function more suitable for the data set,which can further reduce some false detection and missed detection problems caused by threshold setting.These algorithms improve the accuracy of anomaly detection results by 2.1%～3.1%compared with similar methods.3.A video anomaly detection system based on deep learning is realized.Based on the video anomaly detection algorithm proposed in this thesis,the input video sequence is intelligently analyzed,the abnormal frame is detected,and the abnormal result is output.