Research On The Feasible Driving Area Of Narrow Road Meeting Scence Based On Machine Vision

With the rapidly development of machine vision,so analyzing and understanding the target information of in the scene has naturally become a research hotspot.At present,most researchers focus on structured and single simple scene information.,But in real life,smart cars usually need to deal with more complex road scenes,such as meeting scenes,overtaking scenes,and driving scenes with blurred road boundaries and uneven roads.How to accurately understand and explore the content of the scene in real-time under complex scenes to achieve the detection of the driving area of the smart car in the scene is a challenging problem.Thesis takes the meeting scene under narrow roads as an example.Based on the machine vision to research and analyze the width of the road,the driving status of the oncoming car,and identify the meeting area in the scence,so as to detect the drivable area in the scene.Providing accurate and realtime scene information for subsequent smart car path planning or collision avoidance operations.The main work is as follows:Firstly,the feasible driving area detection task needs to analyze the target information in the scene,so the system must have the ability to obtain target information quickly and accurately.Thesis based on the encoding-decoding ideas of SSD target detection algorithm and Deeplab v3 segmentation algorithm,designed a multi-task joint method that shares Res Net-50 feature extraction network.This method can complete the target detection task and segmentation task in the scene at the same time.It guarantees high real-time performance and accuracy and improves calculation efficiency.At the same time,the amount of network operation parameters is reduced and the occupation of computer resources is reduced.Laid the foundation for the efficient execution of other subsequent computing tasks.Secondly,Aiming at the problem of road width detection in the scene,based on the road segmentation information obtained by the joint method,fully consider the position,length,direction and other features of road edge contour edge information extracted by using Canny edge operator,extracted the left and right boundaries of the road,and used the least square method to fit the boundary;Using the principle of inverse perspective projection transformation of spatial road information,a road width calculation model is established,and the detection rules of road width are set,through combining the prior knowledge of road width.Thirdly,aiming at the problem of oncoming vehicle driving behavior and meeting area identification in the scene,combining the target detection frame information obtained by the joint method and the position relationship of the road centerline,the detection and tracking area of the oncoming vehicle is limited.Using the rate of change of the distance between the oncoming vehicle and the road centerline to describe the driving behavior of the vehicle.By analyzing the characteristics of the road surface including the meeting area,thesis proposes a method for judging the traffic area based on road branch analysis.At the same time,Combined with the established longitudinal ranging model,the calculation of the longitudinal distance between the vehicle and the meeting area in the scene is guaranteed.Finally,comprehensive analysis of the obtained road width,oncoming vehicle driving behavior,meeting area,etc,and proposed the detection process of the feasible area in the scene;Then,through real vehicle tests to verify the feasibility of the detection system.