Research On Vehicle Detection And Tracking For Monocular Vision Based On Deep Learning

With the frequent traffic accidents,a series of safety problems caused by traffic accidents have drawn widespread public attention.Faced with the increasingly serious traffic safety situation,the development of the advanced driver assistant system has become an urgent requirement of the automotive industry.The forward collision warning system is an important part of the advanced driver assistant system.Vehicle detection and tracking algorithms play a crucial role in this system.The visual system can get rich content from the surrounding environment,so using visual system for vehicle detection and tracking has been a research hotspot of researchers.In recent years,deep learning develops rapidly.This dissertation is devoted to monocular vision vehicle detection and tracking algorithms research,and designs a set of vehicle forward collision warning system.Firstly,the image is preprocessed,and the edge detection algorithm based on the lane line characteristics is designed to detect the edge of the lane line.The Hough transform is improved by considering the region of interest of the lane line.Then use this algorithm to detect lane lines.The results show that the edge detection algorithm in this paper can effectively filter out interfering objects and detect lane lines more accurately.Secondly,the deep learning algorithm was studied.The video was collected and the training samples were marked.The convolution neural network was designed and trained based on Faster-RCNN.At the same time,the lane line area is used to divide the area of interest of the vehicle.The results showed that the deep learning algorithm can increase the true positive rate,and it still has a good test results in a variety of complex conditions.Thirdly,the Kalman filter tracking algorithm and the Global Nearest Neighbor(GNN)data association algorithm are studied.The Kalman filter tracking algorithm and the global domain nearest neighbor data association algorithm are applied to the vehicle tracking of the video sequence.The results show that the vehicle tracking algorithm in this paper can further improve test results.Finally,the parameters of the camera are calibrated,and the conversion of the image space and the actual space is completed by using inverse perspective transformation to calculate the actual distance of the image target.The standard PC and the data collected by the camera are used to carry out the design of vehicle forward collision warning system.