Deep Learning Based Lane Markings Detection Algorithm

Lane markings detection is an important part of automatic driving system.The main work of this paper is studying the monocular-vision-based lane markings detection algorithm.Namely,the road image of the front side of the vehicle that we processed is obtained by a monocular-vision camera that located in the front window of the vehicle,and then the main lane markings will be detected by our algorithm accurately and fastly.Because the Lane markings are diverse and the environment that they appear is complex,it is need to build strong and robust feature of lane markings.Traditional features and artificial feature have their limitations that they are insufficient to distinguish the lane markings from the road area.In this paper,we use the deep convolution neural network for feature extraction and classification.We should also concern the running time cost of the algorithm.The traditional features of the lane markings are not strong enough for distinguishing,but it is simple and fast,and enough to express and locate the lane markings.Therefore we start from traditional lane detection flow and join the deep convolutional neural network to distinguish lane-markings elements and non-lane-markings elements more effective;namely our lane markings detection algorithm combining the advantage of traditional lane markings detection process and the deep learning.Our lane markings detection algorithm consists of four sub-algorithm modules:preprocessing module,edge extraction module,edge filtering module,and post-processing module.In preprocessing module,we extract the ROI based on priori knowledge for limiting processing area firstly.Then an edge enhancement method that we proposed is applied to highlight the lane markings edge.In the edge extraction module,we use LSD(Line Segment Detector)to extract edges.In edge filtering module,a line segment classifier based on deep convolution neural network is designed to filter the line segment set locally.And then,we use the vanishing point to filter the results globally.The post-processing module mainly consists of the lane markings set generation method and the main lane markings decision method,which is used to generate the main lane markings.The experimental results show that the proposed algorithm has a low computation time with strong robustness and high detection accuracy.