Research On Lane Detection Based On Deep Learning

With the development of deep learning technology,there have been huge gains in the field of autonomous driving technology.Lane detection technology is an important module of autonomous driving technology,so research into lane detection technology is of great significance.Currently in the field of lane detection,research on lane detection technology based on deep learning is a mainstay of development.The main work are on deep learning based lane detection algorithms in monocular vision conditions.Although the research on convolutional neural networks of lane detection has been greatly improved on deep learning methods,currently existing methods can reach over 90%accuracy in regular scenes,but in some challenging scenes the performance is often very poor,with only 30%-60% accuracy,which does not yet guarantee reliable lane detection.The conventional scenarios refer to structured scenarios,where the road structure is simple and the scene is single,the lane markings are clear and the lanes are easy to detect.Challenging scenarios refers to unstructured road scenarios,i.e.driving scenarios that is not trunk and have complex and variable road scenes.Such scenes are often affected by blurred or even badly worn lane lines,obstacles blocking the normal road,light and other factors,so detection in challenging scenes is much less effective than in regular scenes.Therefore lane detection for unstructured driving scenarios is the focus on this paper’s research problem.The main work and points of innovation are listed below.（1）It is proposed to fuse multiple image features to predict the lane features of the next frame.This paper summarizes the following features by looking at the lane detection task: 1)road lanes are line structures,which are continuous on the road;2)two adjacent frames almost overlap;3)lane lines on the road can be obscured by the shadows of other vehicles,pedestrians and roadside buildings,thus making the lane features incomplete.Based on the above characteristics,this paper proposes the idea of a method based on multiple consecutive frames as the sampling object,and the neural network predicts the feature information of the current frame by learning the features of the previous frames,so as to improve the robustness of the network.（2）The fused frame feature module is designed.Firstly,the feature information on each frame is abstract by using convolutional neural network;secondly,multiple consecutive feature maps are input to the recurrent neural network module to build a model from the spatio-temporal sequence to get the fused features of the current spatio-temporal sequence;finally,the lane segmentation map of the current spatio-temporal sequence is output.（3）A novel hybrid model was constructed,incorporating a semantic segmentation model and a long and short term memory network model.Specifically,the feature information of each image is abstracted by the convolutional block of the semantic segmentation model,and then the feature information of several consecutive images（with time-series properties）is fed into the long-short time memory network model unit for feature learning and path prediction.A lane presence prediction branch is also added to predict the probability of the presence of each lane in the full image,helping to obtain more discriminative features of the segmentation network.（4）The performance of the network built in this paper was trained and validated on the classical datasets Tusimple and CULane.The experimental results showed that the accuracy and F1-Measure was 5.579% and 10.6% higher than Enet-SAD in regular and harsh scenarios respectively.The experimental results show that the presented algorithm achieves the objectives and outperforms the currently available methods of all scenarios,and is able to face robust extraction of lane information on real-world complex road conditions.