Design And Implementation Of Highway Service Area Parking Monitoring System Based On Deep Learning

In recent years,with the explosive growth of traffic vehicles in China,parking spaces in some highway service areas have become increasingly tense,illegal parking incidents have occurred from time to time,and people fell uncomfortable when they stay in highway service areas.On the other hand,the parking spaces of oil tankers,animal husbandry vehicles and ordinary cars need to be classified and managed in the highway service area to meet the prevention and control requirements of the highway service area.Considering that the video-based parking space detection technology requires a higher camera position and highly dependent on scene changes,and at the same time this technology also has insufficient recognition accuracy for small targets,which directly leads to a lower recognition rate of the algorithm when classifying vehicle types and is prone to miss detection and false detection.Therefore,this paper aims at improving the algorithm for the problems of low recognition rate of small targets and imbalance of positive and negative samples in the training process in the one-stage detection model,and on this basis,develops and implements the parking space supervision system in the highway service area.The specific work is as follows:(1)On the basis of COCO public data set,this paper adds the sample images taken by high-altitude cameras in the highway service area,re-labels the types of vehicles according to the parking space management requirements in the highway service area,and proposes a convolution neural network structure with multi-scale features on the basis of Yolov3 network structure,thus effectively improving the problem of low recognition accuracy of the original algorithm for small targets;At the same time,this paper adopts mixup data enhancement training strategy to enhance thegeneralization ability of the model,reduce the memory of the model to false labels,and further improve the detection accuracy of the model.Finally,this paper proposes to use swish activation function in the model to solve the monotonicity of Leak Relu activation function in x negative half axis and non-differentiable at(0,0),the nonlinear expression ability of the model is further improved.(2)Aiming at the imbalance of positive and negative samples in the one-stage object detection algorithm model,this paper designs a parameter adaptive deep learning loss function.The loss function not only makes full use of the weight adjustment mechanism of the focus loss function,but also can adaptively adjust the size of super-parameters in the loss function according to the probability value predicted by the model,so that the model can focus on the training of a small number of difficult positive samples,and the model can effectively improves the problem of imbalance between positive and negative samples in the model.(3)Combined with the above-mentioned theoretical methods,this paper designs and implements a parking space monitoring system for highway service areas.The system can detect the number of remaining parking spaces in real time.At the same time,the system can realize the monitoring of vehicles illegally parked in the area.The system can adapt to the complex application scenarios of parking lots in highway service areas,and the algorithm has good generalization ability to identify parking spaces.Based on the detection threshold of 0.5,the accuracy of m AP has exceeded 90%;The detection speed is 25 frames per second.At present,the system has been successfully applied in a highway service area in Zhe Jiang.