Research On Global Lane Detection Algorithm Based On Convolutional Neural Network

Lane detection is an important part of autonomous and assisted driving,which can effectively guarantee driving safety and reduce driving complexity.Therefore,the research of lane detection algorithm is of great significance.For the vehicle driving in the complex environment,lane detection algorithm should be robust and perform in real-time.Traditional method based on Hand-crafted feature cannot deal with the challenging situations to achieve enough robustness.Deep learning based method have become mainstream.This paper studied a high-precision and fast lane detection algorithm using deep learning techniques.The main contents of this paper are as follows:(1)The position lane regression neural network based on the global features is designed to achieve the rough positioning of the entire lane.In the design of the network input layer,a method of lane enhancement based on the maximum projection is proposed.The feature of the lane is enhanced while the image size is reduced,which can effectively increase the speed of the network.The lane node annotation method and feature extraction network based on ResNet are designed,and two kinds of regression neural network ResNet-17 and ResNet-23 are constructed.The predicted lane of the network reaches the average deviation of 5.13 and the detection speed of 4ms per frame.(2)Segmentation of the detected lane at the global stage to obtain local lane block,two local lane center-of-gravity detection algorithms are designed to achieve accurate correction of the lane position.The first one is to build ResNet-11 to detect the center of gravity of the local lane directly.The second one uses a modified K-means algorithm to segment the local lane and calculates the center of gravity of the local lane.Specifically,ResNet-11 reaches the average deviation of 3.81 and the detection speed of 12ms per frame.The algorithm based on K-means achieves the average deviation degree of 1.96 pixels and the detection speed of 50ms per frame.(3)The lane detection based on feature fusion.Inspired by the idea of feature pyramid networks,we combine the low-resolution high-level-semantic feature layer and the high-resolution low-level-semantic layer together,in which the rich location information of high resolution can improve the detect precision.The high-level semantic information contains abundant global feature and can enhance the robustness of the algorithm.The algorithm achieves the average deviation degree of 3.8 pixels and the detection speed of 8.5ms per frame.The data set of this article is from the set of road scenes in China,which is mainly a self-built expressway data set.The training and testing hardware of the algorithm is GTX-1080Ti,and the software environment is MATLAB.