Research And Application Of Lane Detection Technology Based On Convolutional Neural Network

In recent years,with the rapid development of China’s automobile industry,which brings more and more serious traffic problems,safe travel has become one of the main social issues of daily concern.The combination of artificial intelligence technology and traditional vehicles makes autonomous driving technology possible,and advanced driver assistance systems for vehicles are now the main hot spot for research.Lane detection helps guide vehicles to drive safely and can be applied to advanced driver assistance systems.In real traffic scenarios,lane detection is challenging due to the complexity of the road environment,unpredictable weather,dim or dazzling light,and ambiguous lane lines.The current lane detection method cannot detect multiple lane lines at the same time,cannot detect the type of lane lines,and has the disadvantages of low detection accuracy,single detection scene,and poor real-time.To solve the above problems,this paper applies the image semantic segmentation technique to lane detection and recognition.Firstly,this paper adopts the calibration method of Zhengyou Zhang to calibrate the in-vehicle camera,and the results of calibration experiments show that the internal and external parameters of the in-vehicle camera are accurate,and the distortion correction is good.The lane line images were collected using calibrated in-vehicle cameras,mainly including multiple road environments such as congested road sections,high speed sections,and tunnel sections.Image pre-processing is performed on the collected data,manual image annotation work is performed,and a lane line dataset with 24440 images is formed by image enhancement techniques.Then this paper designs an image semantic segmentation model for lane detection,which consists of an encoding module and a decoding module.In the encoding module,Resnet-50 is used as a feature extraction network to obtain high-level semantic information of lane lines;meanwhile,an improved expanded space pyramid pooling module is used to not only reduce the number of model parameters,but also capture the contextual information in multiple scales.In the decoding module,a network structure with multiple fusion of low-level features is used to gradually recover the resolution of the image by fusing lane line detail information;meanwhile,a self-attention mechanism is added,which can obtain global lane line information and improve the model detection accuracy.Finally,a real-world vehicle validation platform is designed and road experiments are conducted.By conducting validation experiments in the actual road environment and analyzing the data at the same time,the validity and feasibility of the lane line detection model designed in this paper are verified and have some practical application value.