Research On The Key Technology Of Intelligent Video Surveillance Based On Deep Learning

With the development of deep learning technology and the wide popularity of digital surveillance equipment,intelligent video surveillance technology has gradually become an important guarantee for the orderly production and life of today’s society.Intelligent video surveillance technology based on deep learning mainly relies on deep learning models to automate data processing,with high accuracy and excellent robustness,but the model itself is too high computational complexity,resulting in poor real-time processing data and improving the cost of model deployment,which leads to the problem of data processing too centralized,making the system less risk-resistant,to a certain extent hinder the intelligent The popularization of the application of intelligent video surveillance technology.In order to solve the above problems,this study mainly focuses on the intrusion warning system based on intelligent video surveillance technology,and the specific work is as follows:(1)To address the problem of the high computational complexity of most target detection models,a lightweight backbone network algorithm S-GhostNet for convolutional neural networks is proposed.This algorithm is based on the GhostNet algorithm,which uses the Ghost Module structure to reduce the computational complexity of the convolution operation and uses an improved residual-like structure to optimize the gradient back-propagation of the depth convolution to improve the accuracy of the algorithm.In the image classification task,S-GhostNet improves the accuracy of Top-5 by 1.9%,1.6%,and 0.8%,respectively,compared with MobileNetV2,ShuffleNetV2,and GhostNet,with similar computational effort.In the target detection task,a one-stage target detection model YOLOv3-tiny and a two-stage target detection model Faster R-CNN were used as the basis for control experiments,and the accuracy of S-GhostNet was improved by 0.3%～1.0% and 0.1%～0.6% respectively compared to the other algorithms,and the computational complexity was significantly reduced compared to the original model.(2)A lightweight target detection model S-NanoDet is proposed to address the problems of poor detection real-time performance and high deployment cost of most target detection models.The model is improved based on the lightweight target detection model NanoDet.The backbone network uses the S-GhostNet algorithm to reduce the computational complexity of the model further while using an optimized loss function to improve the detection performance of the model,and finally introduces a normalization technique CNSN with stronger generalization capability to improve the robustness of the model.In the target detection task,S-NanoDet has lower floating-point computation compared to the NanoDet model,and the average accuracy is improved to some extent in all input dimensions.With its low computational complexity,the S-NanoDet model is suitable for deployment in edge computing devices like smart video surveillance cameras to decentralize data collection as well as the processing and improve the risk resistance of the system.(3)To further improve the detection robustness of the model in surveillance scenarios and expand the application scenarios of the model,an image style migration method based on the improved CycleGAN model is proposed,which can expand the existing dataset images to specific styles at low cost,thus improving the detection accuracy of the model in specific scenarios and enhancing the robustness of the model.In the experiments with infrared image style expansion,the average accuracy of the model detecting infrared images is improved by 3.9% by migrating the expanded dataset.