Research On Lane-line Detection Technology For Unmanned Vehicles

Accurate perception of lane line information is one of the basic requirements of unmanned technology,which is related to the positioning of the vehicle and the determination of the forward direction.Traditional lane line detection methods are computationally intensive,with poor generalization capability and poor real-time performance;the current mainstream lane line detection algorithms are implemented based on deep learning methods,but it is still difficult to meet the needs of complex road scenarios in terms of real-time performance,accuracy and robustness.To address the above problems,this thesis proposes a lane line detection method based on an improved PINet（Point Instance Network）model to achieve accurate detection of the lane line information in complex highway scenes.The work in this thesis includes the following three main aspects.（1）A new lane line labeling method is proposed and a related sample labeling is completed.The dotted-or-solid attribute classification network is designed by using the deep learning method to learn the features of lane lines and perform the dotted-or-solid classification on the test data with an accuracy of 99.38%.The annotation of the data in the lane line detection task dataset with the information of dotted-or-solid attributes is implemented.The annotation method of vanishing point information for different scenes is proposed for the data in the lane line detection task dataset.（2）This thesis proposes a lane line detection algorithm based on the improved PINet network for the problems that the PINet network is not sensitive to the classification of distant lane lines and the accuracy of lane line recognition in complex scenes.The experimental results show that the detection accuracy of the improved network structure is 93.77%for Accuracy and 90.44%for F₁Score,which are 1.53%and 5.06%higher than those of the PINet model,and the error detection rate and false detection rate are 7.61%and 2.63%lower,respectively.（3）Multi-level constraints are added to improve the network perception of lane lines based on the original algorithm branches.The perception of the features around the lane lines is improved by using the prediction of the lane line dotted-or-solid attributes;the image is converted to a bird’s eye view to reconstruct the parallel structure between the lane lines;and the prediction of the vanishing points is used to focus on the image hierarchy.The test results of the improved model in this thesis on real high-speed road data meet both accuracy（Accuracy of 96.73%and F₁Score of 93.98%）and real-time（30+FPS）requirements,and it has been actually tested on the highway and performed stably.