| For a long time,hardware upgrading is the mainstream method adopted by most field range monitoring systems to improve their performance.But only relying on upgrading hardware can not adapt to the future intelligent development of field range video surveillance needs,software and algorithm problems are becoming more and more prominent.The current field range monitoring system has low detection and recognition rate,low classification accuracy,low intelligence,and other problems.In recent years,the prevalence of deep learning has made video surveillance start to transition to intelligence,attracting more and more researchers to research in the field of machine vision and obtaining a large number of detection algorithm results.This paper conducts research on target intrusion detection systems in the field range environment.Its function is to appear as the target of intrusion,quickly trigger the recognition,and at the same time upload the recognized target image to the information management platform to realize the classification alarm and maintain the safety of the field range.How to improve the detection speed and accuracy of the target of the invasion range in the field environment is one of the most concerning issues of the current field range intelligent monitoring system.When the target detection algorithm deals with two different demands of image classification and localization at the same time,the recognition ability of the detection algorithm is not capable of handling the accurate classification task.When dealing with all kinds of targets invading the range in a complex environment in the field,targets with more similar morphology can seriously reduce the classification accuracy of the YOLOv4 algorithm.According to the above problems,this paper focuses on the following research.(1)An improved YOLOv4 algorithm based on Mobile Netv2 is proposed.To address the problem of the low speed of YOLOv4 for target detection in the field environment,a lightweight algorithm model is introduced to replace the backbone network of the original YOLOv4,and a Mobile Netv2-YOLOv4-based algorithm is proposed to increase the detection efficiency of YOLOv4.For the problem that the detection accuracy of YOLOv4 decreases after replacing the backbone network,an improved SPP module and an improved loss function is proposed,and finally,the improved YOLOv4 can quickly detect intrusion targets in the field range environment.(2)A new CBAM-Res Net50(Convolutional Block Attention Module Residual Network)algorithm is proposed.The improved YOLOv4 detection algorithm suffers from the classification performance in the field environment,so this paper adopts Res Net50 as the research object,and first improves the residual block structure in Resnet50 to accelerate the model classification efficiency.Then the top-level features are consolidated and strengthened by the bottom-level features,and the channel attention mechanism and spatial attention mechanism are added to improve the recognition of effective features of specific targets and weaken other interfering features so that the classification algorithm can handle the classification of targets in the field range with richer detail information and faster efficiency to classify targets with close shape,and finally,the improved Resnet50 classification capability is more accurate and efficient.(3)Develop Linux-based field range video surveillance system software and build Raspberry Pi 3B experimental platform.In the actual field range environment,the test equipment verifies the efficiency and accuracy of the improved algorithm proposed in this paper. |
