Research On Video Anomalous Behavior Detection Based On Deep Learning

Smart city is now the main direction of China’s urban development.Urban security and public safety issues are important parts of the smart city.In public places,it is wasteful of resources and inefficient to rely only on human to detect abnormal behaviors in video surveillance.With the rapid development of deep learning,the analysis of video using deep learning techniques has become a new research hotspot.In this paper,three anomaly detection algorithms are proposed for the task of video anomaly behavior detection: future frame prediction for video anomaly detection based on saliency perception,pose clustering based on graph convolution for video anomalous behavior detection,and hybrid future frame prediction and memory reconstruction for video anomalous behavior detection.The specific work is as follows:(1)An algorithm called future frame prediction for video anomalous behavior detection based on saliency perception is proposed.To address the problems of dynamicstatic imbalance and foreground-background imbalance in current methods,this paper uses GAN to generate predicted future frames of the video.In the generator,a modified U-Net is used as the generator,and then a motion-aware module is added to the bottom layer of the U-Net to assign greater loss to the motion part where anomalies occur.In addition,the saliency perception prediction branch is added to extract the saliency features of the video.The saliency features is used to simulate the human eye’s perception of the environment so as to attenuate the foreground-background imbalance problem.In the network optimization section,saliency-guided appearance loss and saliency prediction loss are proposed to optimize the network and achieve real-time video anomalous behavior detection.(2)An algorithm called pose clustering based on graph convolution for video anomalous behavior detection is proposed.For the problems that prediction-based methods are easily affected by factors such as illumination and occlusion,and that most abnormal behaviors in videos are related to people,this paper uses a skeleton model that can effectively solve these problems.The skeleton model can be represented as a spatio-temporal graph over the video sequence,and the pose estimation method is used to extract the key points of each person in each frame,while the appearance features are extracted for the input continuous frames as a way to compensate for the missing feature information in the pose estimation.Then a linear combination of a graph convolutional Autoencoder and a clustering branch is used to map the training samples into a latent space.Finally a Dirichlet process mixture model is used to calculate the score of each sample in the space,and anomalies are detected based on the score.(3)An algorithm called hybrid future frame prediction and memory reconstruction for video anomalous behavior detection is proposed.Since the prediction-based method is vulnerable to the interference of environmental factors in complex scenes and the reconstruction-based method is prone to the problem of overcomplete,this paper proposes a new hybrid video anomaly detection method by combining prediction and reconstruction.The future frame prediction module is used as the prediction branch to make predictions,and then the memory module is used to reconstruct.In order to solve the overcomplete problem,the encoder of the reconstruct module is trained by contrastive learning.The memory module reduces the reconstruction of exceptions by documenting and updating the prototype of the normal pattern.Finally,anomalies are detected by comparing the hybrid frames with ground truth.