Research And Implementation Of Algorithms Of Vehicle Detection In IoV Scene

The rapid increase of vehicles in our country has brought about great traffic problems.The Internet of Vehicles technology is a good way to solve such problems.Thesis mainly studies the roadside unit module in the Internet of Vehicles.Thesis mainly implements the vehicle type detection algorithm.The hardware system is composed of Raspberry Pi4 Model B and Neural Compute Stick 2（NCS2）,and then the vehicle type detection algorithm is transplanted to the built hardware system.On the basis of this algorithm,the number of vehicle types in the traffic environment is counted,so that the vehicle type detection system in thesis is formed.Improvements to the flaws of the dataset.Enhance the dataset required in thesis.Splicing some pictures in the original data set UA＿DETREAC with some pictures in the vehicle type data set,using the generative adversarial network to expand the spliced pictures to generate more pictures with different angles,and finally comparing the newly generated data set with the original data set to form UA＿DETRAC＿CAR.Aiming at the shortcomings of YOLOv3 in vehicle type detection,this paper adopts some measures to improve the vehicle type detection algorithm.Add the Drop Block module to the backbone network of YOLOv3,Use K-means algorithm to re-cluster the new dataset UA＿DETRAC＿CAR to get new anchor boxes,improved multi-scale prediction and training for the actual use environment of the vehicle type detection algorithm,preserving the scale for large and medium-sized objects.The AP and m AP of the improved vehicle type detection algorithm on the new data set UA＿DETRAC＿CAR are basically the same as those of YOLOv3,and the m AP is increased by 0.29%,but the algorithm has been simplified,and the network volume is smaller than that of YOLOv3.When the algorithm is tested on the RTX2060 GPU,the FPS reaches 26.47,and the processing of a single image can reach 54.78 FPS.In cloudy,sunny,rainy and night weather conditions,the algorithm performs well in sunny and cloudy days,and the average detection accuracy of each vehicle model is over 90%.When detecting from the front and rear of the car from a top-down angle,the average detection accuracy of each model has a good performance.In order to make the vehicle type detection algorithm better used in the actual production environment,the algorithm needs to be transplanted into embedded devices.Thesis uses Raspberry Pi,neural computing stick and webcam to form a hardware system.The system is small in size and low in power consumption.In addition,the neural computing power stick can provide a certain computing power support.Such a hardware system is very suitable for deployment in the roadside unit of the Internet of Vehicles.The vehicle type detection algorithm is transplanted into the hardware system and the number of vehicle types is counted in it,and then the video collected in real time by the camera is detected and the number of vehicle types is counted,so as to achieve real-time monitoring of the traffic flow in the traffic environment.On this hardware system,the FPS can reach 10.32 when the video captured by the camera is detected in real time.