Research On Vehicle Object Detection And Ranging Algorithms Based On Monocular Vision And Embedded System

Under the background of the continuous development of 5G and artificial intelligence and the people’s continuous pursuit of a better life,the development of intelligent transportation system with automatic driving technology is an inevitable trend,but automatic driving technology is a huge system science,and the realization of automatic driving is not completed in one step.One of the most important tasks is to know the road condition information and get the location and distance of the vehicle in front.Due to the change and difference of driving scene,vehicle object detection and ranging based monocular vision is still uncertain,and embedded platform has the contradiction of real-time and accuracy in deep learning operation.For vehicle safety,improving the accuracy and efficiency of vehicle object detection and ranging is a long-term problem and challenge.This paper mainly focuses on the vehicle object detection and ranging,studies the vehicle object detection algorithm based on deep learning and the vehicle object ranging algorithm based on monocular vision,gives the corresponding improved algorithm,and studies the codec on the embedded hardware platform.The ultimate goal is to realize the vehicle object detection and ranging algorithm in the embedded system to accurately detect the vehicle,Accurate calculation of the distance between cars,smooth screen playback.The main work is divided into the following three parts:(1)Vehicle object detection.On the basis of theoretical analysis and experimental verification of general object detection algorithm,in order to meet the requirements of embedded platform,the one-step YOLOv3-tiny algorithm based on deep learning is selected as the basic algorithm.An upper sampling layer is added to improve the detection of small object vehicles;The residual network is added to improve the detection accuracy;Re-clustering anchor to enhance correlation and optimize training;The loss function is improved to reduce the training parameters and simplify the training process.The improved E-YOLOv3-Vehicle-Res vehicle object detection algorithm on the embedded platform TX2 has a detection speed of 26 FPS and an average precision(AP)of 83.9%.Compared with YOLOv3-tiny,it improves the detection accuracy on the premise of ensuring the efficiency,and is suitable for the embedded system platform.(2)Vehicle ranging.Based on the monocular vision geometric ranging model,the error causes of the ranging model are analyzed.By using the road vanishing point detection method,the attitude angle is estimated and the monocular vision geometric ranging error is corrected to solve the influence of the camera attitude angle change on the ranging.In a certain range,the relative error of the improved monocular vision geometric ranging algorithm is controlled within 7%,which is nearly 50% lower than that of the original ranging method.It can effectively reduce the ranging error and is suitable for vehicle platform.(3)Embedded system porting and hard coding.Using NVIDIA Jetson TX2 platform to complete the hardware and software design,the transplantation of E-YOLOv3-VehicleRes vehicle object detection algorithm and monocular vision geometric vehicle ranging algorithm to correct the camera attitude angle is realized;The hard coding,push-pull stream and soft decoding based on TX2 are developed.The delay of each video frame is reduced to 43 ms,and the delay of soft coding is more than 200 ms,which solves the problem of video image jamming and greatly reduces the delay.In the real vehicle experiment,the whole system accurately detects the vehicle object in front and selects the position by frame;Accurate measurement of vehicle distance;The embedded platform can play the real-time processing picture smoothly without obvious jam and delay.To a certain extent,the research of vehicle object detection and ranging based on monocular vision and embedded system is completed...