Research On Lane Detection Algorithm Based On Deep Learning

In recent years,unmanned driving technology has become a hot research field in various countries,and lane line detection algorithms,as an important part of unmanned driving systems,still have many problems worth studying.This paper mainly studies the lane line detection algorithm based on the deep learning monocular visual image,that is,the deep learning algorithm is used to process the picture of the road environment information in front of the unmanned vehicle collected by a single camera,and the position and position of the lane line are quickly and robustly obtained.Along the extension direction.There are many problems in the lane line detection task.First of all,the lane line itself is different from ordinary objects.It has a slender characteristic and a large spatial position.Therefore,it is extremely vulnerable to occlusion and causes spatial discontinuities.In addition,the lane line will gradually fade with the wind and sun,and it will become blurry and difficult to recognize.The road environment where the lane line is located is very complicated,and light,water stains,dust,shadow,etc.will all affect the vision-based lane line detection.Therefore,traditional lane line detection algorithms cannot achieve good detection results.Deep learning algorithms show strong capabilities in various fields,so this paper uses deep convolutional neural networks to detect lane lines.In this paper,according to the characteristics of lane line objects,a deep convolutional neural network is used to semantically segment the lane lines to distinguish whether each pixel belongs to the lane line,and then to combine the lane line post-processing algorithm to solve the curve parameters of each lane line.The lane detection algorithm proposed in this paper is roughly divided into two parts: the lane segmentation semantic segmentation neural network and the lane line post-processing algorithm.The overall design considers both the detection effect and the algorithm real-time.The neural network model of lane line semantic segmentation designed in this paper draws on the structural characteristics of the current excellent semantic segmentation model.It adopts an encoder-decoder structure and uses a realtime classification network as a backbone network for fine-tuning.In view of the large spatial distribution of lane lines,the need for The network has a strong ability to extract global features.It introduces advanced self-attention mechanisms in the field of deeplearning.For the efficiency of algorithm operation,an improved channel self-attention module is proposed,which greatly reduces the computational cost of the original selfattention module.At the same time,the linear feature enhancement module is proposed to improve the continuity of the lane line segmentation results in view of the slender shape characteristics of the lane line.Finally,the network output part is divided into two-class output and multi-class output.For the multi-class network,adding a lane line presence prediction branch can improve the robustness of the multi-class network prediction results.The lane line post-processing algorithm mainly includes lane line clustering and lane line curve fitting.In order to improve the efficiency of the post-processing algorithm,the lane line segmentation mask is first sampled,and then different processing is performed on different network outputs.For the binary segmentation semantic segmentation network,the DBSCAN clustering algorithm is used for clustering,and then the lane line is curvefitted using the RANSAC algorithm.For multi-classified semantic segmentation networks,curve fitting is performed between them.By comparing the results of the two semantic segmentation networks,the lane line detection algorithm which is fitted by the multiclass semantic segmentation network combined with the RANSAC algorithm is finally adopted.Experiments show that the lane detection algorithm proposed in this paper has strong detection effect and robustness,and it can cope well even if the lane line is blocked.At the same time,the algorithm has a faster detection speed.On the Nvidia Tesla P100 GPU,this paper is designed The semantic segmentation network can achieve a segmentation speed of 63 fps,and the entire lane line detection algorithm can achieve a detection speed of 40 fps to meet real-time requirements.