Research On Lane Line Detection Algorithm Based On Convolutional Neural Network

As an important guiding sign for safe driving of cars,lane lines are the necessary data collection objects for on-board sensors and the basis for vehicles to achieve autonomous driving.With the rapid development of automobile industry technology,autonomous driving technology has higher requirements for lane line detection technology.It must be able to accurately detect lane lines in a variety of complex scenarios,and achieve all-weather detection with strong real-time and high accuracy.Traditional lane line detection algorithms are generally based on artificial experience design features.The algorithm has high requirements for experience,and has poor ability to deal with complex environments,and its robustness and accuracy are not ideal.In recent years,deep learning methods have developed rapidly and have been well applied in the field of machine vision,solving many problems.Therefore,more and more people hope to use deep learning methods to improve the robustness of lane line detection results.and accuracy.Aiming at the problems of poor robustness and accuracy of traditional algorithms,this paper adopts an algorithm based on convolutional neural network(CNN)combined with post-processing to quickly and stably detect lane lines in a variety of complex scenarios.The specific research work is reflected in The following aspects:(1)Collect lane line images for different road conditions and scenes,classify and organize these images according to the scene,and make a self-built lane line data set,and then use three image preprocessing methods to preprocess the data set to achieve The image data is enhanced and the unnecessary image background is reduced,and then the preprocessed lane line images are marked by the labeling software Labelme,which is then made into a training set.(2)A lane line detection method based on global image feature extraction based on row(Row)is proposed.This method divides the image into several grid cells(cells),and the detection process is regarded as selecting a certain Row on a predefined Row.These cells are selected and classified based on the lane position in the row direction.Then,the CNN model is built through the deep learning framework Pytorch,and the VGG16 convolutional neural network is used as the backbone network to extract the lane line features.Then an auxiliary branch is also designed,and an auxiliary segmentation method using multi-scale features is proposed to extract local features.Finally,a loss function of classification loss combined with structural loss is designed to make full use of the prior information of lane lines,such as the physical structure and continuity features of lane lines,and use them as constraints of network output results.Row-based feature extraction The method significantly solves the two detection problems of high computational cost and no visual information.(3)The post-processing operation is used to process the pixel embedded image of the lane line segmented by the network model.First,the lane line pixels divided by the network model are clustered by the clustering algorithm,and then the clustered lane line pixels are fitted by the least square method,and the positions of each lane line are connected according to the clustering points.The fitted lane line pixels are returned to the original lane line image,and the lane line pixel categories are completely divided through post-processing operations to avoid the interference caused by some unnecessary pixels.(4)Through the ablation study on the Tusimple dataset,the effectiveness of the grid unit and each module is verified.The experimental results show that the optimal detection accuracy can be obtained by setting the number of grid cells to 100,and the network performance can be improved by adding lane structure loss and feature aggregation.Then it was tested on the Tusimple public lane line data set and the self-built lane line data set.The test results show that the accuracy and detection speed of the algorithm in this paper are better than several other typical CNN algorithms.The test on the Tusimple public data set The accuracy rate reached97.06%,and the test accuracy rate reached 91.3% in the ordinary scene of the self-built data set.The detection speed can reach 175.4 frames per second,which meets the real-time detection requirements and verifies the CNN-based lane line detection proposed in this paper.Efficiency and robustness of the algorithm.The CNN-based lane line detection algorithm proposed in this paper has a good ability to deal with complex road scenes,and has a certain theoretical reference value for driver assistance systems and driverless perception.