Research On The Lane Detection Method Based On Deep Learning

Lane detection is one of the basic tasks of automatic driving,which is used to regulate and guide vehicle driving,and has important research significance.The traditional algorithm uses features such as geometry and color to detect lanes.The algorithm is simple,but it cannot cope with complex traffic scenes.In recent years,lane detection has gradually developed into the field of deep learning.In particular,algorithms based on semantic segmentation have been widely used.The accuracy and robustness of lane detection algorithms have been greatly improved,but there are some problems such as slow computation speed,limited receptive field of pixels and low detection efficiency.To solve these problems,a lane detection method based on deep learning is proposed.The specific content is as follows:(1)Multi-branch feature network design based on lane polynomialA multi-branch feature network based on the lane polynomial is designed.The lane is regarded as a curve polynomial.The deep learning model is used to predict the key parameters of the polynomial and locate the lane,which mainly includes the trunk network and the polynomial parameter prediction.The backbone network is designed as a multi-branch structure.The upper branches are improved based on ResNet structure.First,the initial large convolutional kernel is disassembled,and then the convolutional module is added to bottleneck structure to delay down sampling and improve the efficiency of feature extraction.The lower branch of the network is based on the shape priori of the lane,and the slice hierarchical sequence structure is designed.The spatial features between the row and column pixels are fully utilized,and the upper and lower branches are combined to form a multi-branch feature extraction network.(2)Key parameter prediction of lane polynomialAfter feature extraction,the key parameters of the lane polynomial,including polynomial coefficient,the deviation of the longitudinal starting and ending points of the lane,and the transverse deviation of the lane,can be predicted directly by the trunk network to standardize the position of the lane.Then the multi-task loss function is designed and the algorithm model is trained using Tucson dataset.The accuracy of the final algorithm is 95.85% on the Tucson test set,and the detection speed can reach 80 fps.(3)Design of lane post-processing schemeIn order to improve the detection algorithm and save calculation power,the post processing method of searching lane with edge sliding window is designed.The lane detection is used for real-time video processing,and there is continuity between videos,so after using the multi-branch feature network to locate the lane,the edge sliding window can be used to search the lane in the subsequent video.The combination of the two improves the real-time performance of the detection algorithm,and enables the algorithm to run in realtime on the embedded device(NVIDIA-TX2).(4)Lane identification experimentThe model vehicle and simulation map were designed to test the feasibility of the detection method of multi-branch feature network combined with lane post-processing technology.In the experiment,the vehicle model is equipped with NVIDIA-TX2 for real-time image processing,and the lane detection model is generalized to the simulation map data set by means of transfer learning.The results show that the vehicle model can finish driving autonomously,which proves that the lane detection method has high accuracy and real-time performance.