Research And Application Of Target Detection Algorithm For Edge Computing

In the field of modern security protection,manual video monitoring method based on cloud computing is often used.When monitoring abnormal targets with low probability,the monitoring system is prone to the problems such as the waste of network resources due to redundancy of video information,the delay of data transmission and the decline of detection quality caused by human error.Edge computing performs real-time analysis and operation on the collected front-end data at the edge of the network,reduces the cloud load pressure effectively and improves the real-time response of the system.At the same time,with the rapid development of deep learning technology,especially target detection algorithm,video surveillance is more efficient and intelligent,and the defects existing in artificial video surveillance methods are effectively alleviated.Therefore,this paper proposes a scheme for edge computation-oriented video surveillance system,which utilizes the target detection algorithm based on deep convolutional neural network to realize the efficient capture of abnormal targets,and verifies the feasibility of the scheme through the monitoring data of coalbed methane well station.Specific research contents are as follows:Firstly,in view of the performance problems existing in the cloud-based manual video surveillance scheme,a video surveillance system architecture scheme oriented to edge computing is designed.The scheme uses NVIDIA TX2 as the edge device to perform data collection and forward reasoning of detection algorithm.PC is the edge server,which is executes the video data and detects the algorithm parameter file cache.The intelligent cloud server performs algorithm parameter iterative update and regularly pushes parameter files to the edge server to achieve collaborative work.The feasibility of the design scheme is verified by monitoring data in coalbed methane well station.Secondly,a lightweight deep convolutional network named MMGNet for edge computing is proposed to solve the problem that the feature extraction network of the current target detection algorithm has large parameters and high operating memory.The network can effectively reduce the calculation consumption and storage requirements of the model by the depthwise separable convolution,and further realizes the model light weight with Ghost Net.The proposed network maintains the form of "expansion first,then compression" in structure,and introduces channel attention mechanism to improve the model accuracy.In addition,the proposed network also adds multi-scale convolution kernel to extract feature information,which simulates people's view of object characteristics from different perspectives,and further improves the detection accuracy.The TOP-1 accuracy rate of 71.52% is achieved on CIFAR100 dataset.Finally,aiming at the abnormal targets monitoring method in video surveillance,the target detection algorithm based on MMGNet is adopted to realize the rapid and accurate capture of abnormal targets.In this paper,Faster-RCNN,SSD and YOLOV4-tiny are used as the basic algorithm framework,and the feature extraction network is replaced by MMGNet designed in this paper.Then,through comparative experiments on part of Pascal VOC 2007 and 2012 target detection data sets,it is verified that the designed network can effectively reduce the parameters of target detection algorithm,and the detection accuracy and running speed can meet the actual needs.After horizontal comparison of various indicators,YOLOV4-tiny algorithm based on MMGNet is selected to test the manually labeled data set of coalbed methane well station.The average classification accuracy is 92.15% and the detection time of single sheet is 0.102 s when running on the TX2 platform.The test results show that the proposed method is feasible and effective.