Deep Learning Based Multi-module Joint Video Anomaly Detection Research

Video anomaly detection(VAD)has been the focus and difficulty in the field of computer vision research.Video anomaly detection not only needs to be detected with high accuracy,but also needs to improve its inference speed while ensuring accuracy,so that it can meet more low-performance devices.How to reduce the demand for computational resources while guaranteeing a certain level of detection accuracy,so that it can be better applied to different performance levels of computing devices is the problem that needs to be solved nowadays.Based on this goal this paper conducts a research on video anomaly detection algorithm,and the main research contents of the paper are as follows:1)To address the problem that the storage capacity of current neural network models used for video anomaly detection is poor,which leads to easy loss of feature information and reduces the detection accuracy,a study on the storage capacity of neural networks is carried out,and a video anomaly detection network(SMDNet)with Spatio-temporal Memory-guided Dynamic balanced margin learning is designed.Based on Conv LSTM,the spatio-temporal memory module(Conv LSTM-SMM)is proposed to compensate for the lack of storage capacity of the neural network and applied to video anomaly detection based on deep learning framework.The designed spatio-temporal memory module continuously updates the feature parameters in the memory network by querying the features,and stores the updated feature parameters with the query features for matching,so as to obtain new feature parameters to enhance the spatio-temporal features of the network.Meanwhile,in order to ensure the stability of the training network,the dynamic balanced margin learning mode(DBM)is proposed based on margin learning,and the spatio-temporal memory module is used to guide the dynamic balanced margin learning,and different sample spacing is used in different training phases of the samples to help the network get the suitable edge distance in different training periods,so that the detection accuracy can be improved.The experimental validation part is conducted on the most challenging public datasets(Avenue,Shanghai Tech)at present,and the experimental results show that the designed video anomaly detection algorithm with spatio-temporal memory-guided margin learning has improved the video anomaly detection rate and superior detection performance.2)The rapid development of deep learning has made deep neural networks an effective tool for solving complex problems,but high accuracy means that the computational complexity of the model is also constantly increasing,In order to achieve lightweight computing,this article designs a Fire Res Net based pruning accelerated video anomaly detection network(FRPNet)to improve the inference speed and performance of video anomaly detection algorithms.Based on the Cycle GAN main framework,a lightweight module Fire module is introduced in the coding section,and the Fire Res Net module(FRM)is designed.Pruning operations are used in the coding layer to reduce the number of model parameters,while maintaining expression and feature extraction capabilities while reducing computational complexity.The pruning algorithm adopts gradient descent training method,which ensures the reduction of model parameters,avoids the risk of overfitting caused by gradient disappearance,and improves the generalization ability of the model.The experimental verification was conducted on the most challenging public datasets currently available(Avenue,Shanghai Tech).The experimental results showed that the designed FRM based pruning acceleration video anomaly detection algorithm ensured the detection rate of video anomalies,while improving computational performance.It can significantly improve the running speed and accuracy of deep learning models on small devices,which is conducive to the practical application and promotion of video anomaly detection.