Research On Intelligent Method For Lane Detection Of Complex Working Conditions

The "Guide to the Standard System Construction of the National Internet of Vehicles Industry(Intelligent Connected Vehicle)" points out that the Intelligent Connected Vehicle is a new industry with deep integration of vehicles,electronics,information and communication,road transportation and other industries,etc.It is a global innovation hotspot and a high point of development in the future.The intelligent connected vehicle is equipped with advanced in-vehicle sensors,controllers,actuators and other devices,and integrates modern communication and network technology to realize vehicles and X(vehicles,road,people,cloud,etc.)exchange and sharing information intelligently,with complex environment awareness,intelligent decision-making,collaborative control and other functions.It can achieve safe,efficient,comfortable,energy-saving driving,and finally realizes a new generation of cars that replace the operation of people.Among the intelligent driving technologies,the environmental awareness is the key to ensure the intelligence and safety of intelligent vehicles.Moreover,the lane detection is the core technology of environmental awareness and premise of safe and autonomous driving.However,it is also the difficulty of research and the core of technology.In this study,the research on the intelligent detection method of lane for complex road conditions is carried out with the support of 2016 Intelligent Manufacturing Integrated Standardization and New Model Application Project "Construction of Research and Test Verification Platform for Intelligent Network Auto System and Communication Standardization"(No: 2016ZXFB06002)of the Ministry of Industry and Information Technology,and Chongqing Science and Technology Plan Project Key Industry Common Key Technology Innovation Special Subject Project "Standardization of ADAS Functional Testing and Construction of Road Test Certification Platform"(No.CSTC 2015zdcyztzx60005).The main research contents and innovations of this paper are as follows:1)Aiming at the problem that training a deep learning model requires a large number of label images,the automatic generation algorithms of front and lateral label images of vehicles are studied.By studying the color space of LUV and LAB,a segmentation algorithm for yellow and white lanes in simple scenes is proposed,which provides a large number of label images for the training of the forward lane detection model.For the lateral lane label images,the strategy of synthesizing a large number of label images using a small number of images in real scene is been designed.Then the Image Quilting algorithm is been improved using convolution neural network,the quality of synthetic lane and asphalt pavement images are improved significantly,thus the synthesized label images can meet the training requirements of in-depth learning model.2)Aiming at the problem that the detection accuracy of the forward lane is not high in complex scenes,the YOLO v3(S×2S)algorithm is proposed according to the distribution law of the lane in the bird's-eye view.In order to learn the forward lane features in complex scenes automatically,a two-stage lane feature learning framework is designed based on YOLO V3(S×2S).So the lane features in different scenes can be learned adaptively.As the result,the training efficiency of lane detection model is significantly improved,and the forward lane detection in complex scenes is been realized.3)It is difficult to detect lanes under complex conditions such as occlusion,deformation and wear.To solve the problem,the structure relationship of lanes is systematically studied.The strategy of lane position prediction by using the relationship of lanes is proposed.Based on the historical position information of the detected lanes,a lane prediction model,LSTM-BGRU,based on two Bidirectional GRU is prosed,and another model named as ALSTM-RcNN is been proposed using recurrent neural network and the angle of lanes.In order to ensure the accuracy of the final lane detection result,the D-S fusion algorithm is used to fuse the output results of the lane detection model YOLO v3(S×2S)and the lane prediction model ALSTM-RcNN,so that the optimal lane position can be obtained finally.4)For the problem that the distance between the lateral lane and vehicles is difficult to calculate,the distance of the lateral lane is been discretized into several distance bars averagely.Thus the calculation of the distance of the lateral lane is transformed into a multi-object recognition problem.A deep convolutional neural network,LatDisLanes,is been put forward to achieve end-to-end detection of distanced between vehicles and lateral lane.In addition,by analyzing the corresponding relationship between the world and the pixel coordinate system,a dynamic correction model is been proposed to correct the recognition error of LatDisLanes caused by the change of inclination angle in real time,which ensures the recognition accuracy of the distance between the lateral lane and the vehicles.