End-to-end Lane Detection Method Based On Deep Learning

As a basic task in the field of automatic driving,lane detection has been widely studied.It can be divided into traditional computer vision methods and deep learning methods.Most traditional computer methods use edge detection and Hough transform to extract lane lines.It includes a lot of manual engineering,which requires complicated adjustment experiments and is very skillful.This method is limited by the poor robustness of the hand-designed features,and most of them can only show good results for specific scenes or specific data sets.With the rise of deep learning,people began to use deep neural networks to identify lane lines.At present,most of them are image semantic segmentation methods based on FCN(Full Convolutional Neural Network)or target detection algorithms based on YOLO and SSD.These methods are essentially classification methods.The lane lines are found by classification and positioning methods,and the lane lines that are located are converted into line information by post-processing.This paper proposes an end-toend deep learning lane detection model,which is different from the traditional classification algorithm,which uses the regression fitting method to extract the lane lines.The process of extracting lane line extraction and post-processing extracting lane line curve parameter information in the traditional model is integrated,which eliminates the post-processing process of the model and speeds up the network operation speed.At the same time,the traditional classification deep learning method is based on the probability model.When the lane line defect or occlusion occurs,the model can not accurately determine the lane line information of the area.The improved regression depth learning method performs regression fitting on the lane line curve.In the case of local lane line information defect,there is still good detection and detection effect,and the robustness of the model is better.The core content of this paper aims to apply the regression-based deep learning method to the lane recognition task.The main research work of this paper is as follows:(1)A lane line data set dedicated to the regression model was created,and more than 20,000 pictures were obtained by taking a video around Hunan University and intercepting it.A lane line data set dedicated to regression recognition is created by performing image processing,filtering,labeling,data set encapsulation,and the like on the image.Compared with the open source dataset,the dataset not only marks the location of the lane line,but also saves the location information of the laneline to the dataset label in the form of a quadratic equation parameter.(2)Introduce the principle and components of convolutional neural network.By constructing a deep learning model of convolutional neural network and training and testing on the previously constructed data set,the results prove that the regression-based deep learning model can be effectively completed.Lane line detection mission.At the same time,the direction of further optimization is proposed for the training results of the model.(3)In order to further improve the accuracy of model recognition,this paper combines the traditional computer vision technology to propose an image preprocessing module combining RGB color space inspection algorithm,HSL color space detection algorithm and edge detection operator..The module is combined with the deep learning model to carry out the training test.The training test results of the deep learning model with feature extraction module and the deep learning model without feature extraction module are compared.It is proved that the feature rough extraction model is effectively improved based on The recognition accuracy of the regression deep learning model in the lane line detection task.(4)Experiments were carried out on the algorithm model through the smart car platform of the laboratory,mainly testing the real-time and generalization performance of the model.Firstly,the model is migrated to improve the recognition performance of the model on the experimental runway.Then the visual model of the car is connected with the car control system,and the output of the visual model is used as the input of the car steering.The experimental results show that the real-time performance of the model basically meets the driving requirements of the car,and the algorithm still has good performance on the new lane line,which indicates that the algorithm has good generalization performance.