Object Detection Method Based On Deep Learning In IoV Scene

With the rapid development of the automobile industry at home and abroad,the urban traffic pressure is increasing.The Internet of vehicles(IoV)system is an important breakthrough point to improve the urban traffic pressure.The IoV system includes vehicle unit and roadside unit.Roadside unit is responsible for sensing all kinds of road information.As a real-time information driven service system,the information perception ability of roadside unit,which acts as its information source,is very important.This paper will mainly study the information perception of roadside unit,and combine the streetlight and roadside unit as the intelligent streetlight system of the Internet of vehicles.There are many ways of information perception.This paper focuses on the object detection methods based on deep learning to perceive the road traffic information contained in the image captured by intelligent street lights.The main research contents are as follows:First of all,this paper mainly studies the problem of image information perception in the way of deep learning,so we will briefly introduce the related concepts of deep learning,the operation process of neural network,and the related layers of convolution neural network.After that,UA-DETRAC-LITE-NEW data set is introduced,which is a mixup dataset built by us and suitable for application in the scene of IoV.Secondly,an improved retinanet vehicle detector is proposed.The resample module of ResNet-50 can’t make full use of input information,so we.improve its structure,and the channel attention mechanism is added to each residual block.After that,the NMS strategy is optimized.Finally,the mAP is increased by 3.11 percentage points to 89.63%,and the detection speed on RTX2080 is 13.19 FPS.Then,an improved yolov3 vehicle detector with high real-time performance is proposed.The new priori information is clustered under the vehicle detection scene,and the dataset is further enhanced by using of Mixup.The confidence of the model is reduced and the generalization ability of the model is improved by using of Label Smoothing.Afterwards,the position regression loss is improved by using GIoU,the confidence loss is improved by using fine-tuning focal loss,and the network structure is improved by using CBAM module.Finally,the mAP increased by 7.02 percentage points to 88.35%,and the detection speed on RTX2080 is 31.14 FPS.Finally,two improved vehicle detection algorithms are proposed for the intelligent street lamp system of the Internet of vehicles.The first one is a fine-grained traffic flow statistics algorithm suitable for the adaptive virtual coil in the object detection scene.The specific algorithm design is given and the performance of the algorithm is verified by experiments.The second is the intelligent intersection signal light control system,the function design and the signal light control algorithm design are given,and we did simulation experiment in the Guiyang actual traffic flow data.The signal light control algorithm can reduce the waiting time for vehicles to pass through the intersection to a large extent in the peak period.