Research On Target Detection And Forward Collision Warning System Based On Deep Learning

In recent years,the steady growth of China’s economic level has led to a rapid increase in domestic car ownership.Although the car is convenient for the daily travel needs of residents and has become the most widely used means of transportation,it also makes traffic accidents occur frequently,and the driving safety problem has attracted more and more attention.The research of advanced driver assistance system has become a hot spot.It is the development trend to apply the deep learning method with excellent performance in the field of computer vision to the direction of automobile driver assistance.The topic of this paper is to study the front collision warning system based on deep learning.The research object is the vehicle and pedestrian targets in front of the road.The core content of the research includes the target detection algorithm,target tracking algorithm and distance measurement algorithm based on monocular vision.The forward collision warning system inputs the front road image into the target detection and target tracking model to realize the target tracking in the road environment.The monocular vision ranging method is used to calculate the distance and position information of the road target in front.Finally,the speed of the road target is analyzed according to the position information of the target to judge whether collision warning is needed.The main work of this paper is as follows:1.A target detection network model suitable for the forward collision warning system is constructed.In order to make the target detection model meet the performance requirements and have the advantages of lightweight and good real-time performance,this paper improves the fusion of Mobile Net V2 network and YOLOv5 s network.The improved model uses the channel pruning method to get the YOLOFCW network model,and finally uses the knowledge distillation training method to restore its accuracy.2.Based on the target detection model,the target tracking model Deep SORT is constructed,and the feature extraction network of the target tracking model is improved.Aiming at the problems of feature extraction network in Deep SORT,four kinds of feature extraction networks with different sizes are designed to meet different development needs and further improve the speed of target tracking model.At the same time,it solves the problems of the sudden disappearance and violent shaking of the boundary box detected by the YOLO-FCW model,so as to realize the real-time tracking of the detected target.3.The distance detection algorithm of road target in front is constructed.Aiming at the problems of monocular vision,a compound ranging algorithm based on lane detection is proposed.Firstly,the algorithm detects the lane line and calculates its curvature.According to the horizontal pixel width between the lane lines at the position of the front target in the two-dimensional image,the real pixel width is calculated by using the triangular relationship,so as to complete the ranging work.When there is no lane line in the road,the distance measurement method based on projection model is used.4.The model of front collision warning system is constructed.Through the analysis of the braking distance of the vehicle,the first level safety distance and the second level safety distance of the front collision warning system are obtained.Combined with the detected target types,the road ahead is divided into level 1 and level 2 collision warning areas to realize the multi-level early warning design of the early warning system.