Research On Vehicle Inspection Method Based On Deep Convolutional Neural Network

In recent years,China’s automobile industry has undergone rapid development,and the traffic safety situation is serious,and violations of the traffic behavior also occur from time to time.For a better traffic environment,the traffic system needs to be intelligent,combining pedestrians,vehicles,and roads to run these traffic information to solve the problems in traffic.In the intelligent traffic system,vehicle detection is a critical component of the system,and the given image information is used to determine whether the image contains a vehicle,and to determine the location of the vehicle.This thesis focuses on the detection of vehicle information in video streams by means of deep learning.For YOLOv5 has the drawback that the model recognition accuracy is not high enough under the vehicle detection task in complex traffic scenes.In this thesis,the YOLOv5 algorithm is improved by adding three attention mechanisms,SE,CBAM and CA,respectively,to suppress the complex background information of vehicle detection,so that the model can pay attention to the vehicle information during the detection.By comparing the improved detection accuracy and real-time performance,CA attention is finally selected,and the mAP of the improved model is 98.2%,which is a 1.1%improvement,and the FPS on RTX2060 is 73.5.To address the shortcomings of the YOLOv5 vehicle detection algorithm with added CA attention in terms of real-time performance and model size,this thesis improves it in the following three aspects: optimization of prior information,improvement of multiscale prediction,and improvement of feature extraction network.To address the use scenario of the vehicle detection algorithm,the original prior frame information is not well adapted to the task scenario of vehicle detection.First,the vehicle data set is reclustered to obtain new prior information,and then the large target scale prediction is cropped,and the scales of medium-sized targets and small targets are retained to reduce the size of the model.After that,the RepVGG module is fused into the feature extraction network in order to further improve the real-time performance of the algorithm.The final mAP of the improved YOLOv5 vehicle detection algorithm is 98.45%,an improvement of 0.25%,and the FPS on the RTX2060 reaches 109.9.Finally,an improved vehicle detection algorithm is proposed for the application scheme of vehicle flow statistics system.The traffic flow statistics system can use the vehicle detection algorithm to detect vehicles on video,and realize the real-time statistics of traffic flow by SORT algorithm.This thesis adopts the method of virtual detection line to achieve the function of traffic flow statistics,by comparing the counting results of virtual detection line in different locations to find the appropriate location of the detection line,the experiment found that the location of the selected virtual detection line in onehalf of the entire video,the statistical results of the traffic flow statistics system is closer to the real results,the accuracy rate of up to 93.8%,thus proving that the improved vehicle detection algorithm has practical engineering value.