Research On Front Vehicle Collision Warning System Based On Deep Learning

With the rapid development of current traffic,the number of vehicles and drivers is increasing,and the following problems such as distraction of drivers are causing more and more traffic accidents.In order to reduce traffic accidents,thesis proposes a front vehicle collision warning system based on deep learning.In this system,an on-board camera is used to obtain time-stamped image data in front of the vehicle,and then a deep learning algorithm is used to detect objects in the image.The fusion distance measurement method is used to calculate the position information of the vehicle in front.Then combine the distance and time information to analyze the vehicle speed,and finally carry out a comprehensive early warning: use TTC warning model for vehicle collision warning,and use the distance between non-motor vehicles and pedestrians to carry out early warning of non-motor vehicles and pedestrians.The main research work of thesis consists of the following three parts:1.Propose a target detection algorithm that is suitable for traffic scenarios which can be used in low-cost embedded platforms.The target detection algorithm is the first step of the early warning algorithm,whose accuracy determines the performance of the early warning system.thesis analyzes the geometric scale features of traffic targets,then designs the network,and uses the residual structure to connect the bottom-level features and high-level features,and performs target detection on two scales.Experiments show that this algorithm can ensure the accuracy of detection while performing target detection quickly.2.A fusion distance measurement method combining distance measurement based on license plate size and distance measurement based on geometric model is proposed.Analyze the advantages and disadvantages of license plate size ranging and geometric model ranging through experiments,it is found that the method based on license plate ranging is prone to the problem of ranging failure caused by the failure of license plate detection when the vehicle body is far away.The geometric model ranging method is prone to problems such as inaccurate positioning because the landing site is blocked.In order to improve the robustness of the ranging method,a fusion distance measurement method is proposed.The geometric ranging model is first used to roughly divide thevehicle body into a short distance or a long distance.Different ranging methods are used to improve the robustness of the ranging method Great.3.Constructed a comprehensive early warning model including early vehicle collision warning and non-motor vehicle and pedestrian approach warning.Each time the target detection results and target ranging results are saved in combination with the time stamp,and collision warning are performed.Because the behavior of non-motorized vehicles and pedestrians is unpredictable,We set up a collision warning method for non-motorized vehicles and pedestrians.When it is detected that non-motorized vehicles and pedestrians are present in the warning range,early warning is performed.