Deep Learning Lane Line Detection Algorithm Research

The current smart car research on assisted driving is divided into two main types of approaches.One type of approach is Tesla Technology.The logic of Tesla Technology is to imitate the vision system of the human eye by applying neural net operations to the images captured by the camera to analyze and calculate them,and directly derive the results without the help of other perceptual information,using only vision to achieve autonomous driving.Another type of approach is the vision and lidar fusion inspection,which is used by most Chinese automotive companies.Either way,lane line detection is one of the main techniques.Therefore lane line detection has received a lot of attention from researchers and is one of the important research areas in computer vision.Traditional lane line detection is achieved with the help of image pre-processing means,such as edge feature segmentation and feature extraction.However,traditional methods are susceptible to factors such as ambient lighting,vehicle movement,and road degradation.With the development of deep learning,especially the advantages of convolutional neural networks in image tasks,lane line detection methods have become more diverse.Deep learning-based lane line detection models often have large architectures,leading to problems such as long network inference time,too many parameters to be trained leading to difficult network convergence,and in actual detection scenarios,lane lines can be affected by conditions such as mutilation,breakage,and strong illumination,resulting in poor detection results.In this paper,we design a lane line detection model with high generalization by self-attentive distillation and spatial convolution and other structures on the development and shortcomings of lane line detection algorithms for deep learning in detail.The detailed work is as follows.(1)Several classical deep learning-like lane line detection algorithms are analyzed in detail,and their structures and features are discussed.In order to compare the algorithms in terms of structure,performance,lane detection effect and quantitative metrics,relevant subjective and objective experiments are conducted on several datasets.Lane Net,VPGNet,EL-GAN and SCNN algorithms are evaluated and analyzed on the lane line detection datasets Tu Simple and CULane.By analyzing the advantages and shortcomings of each algorithm,the self-attention distillation mechanism is applied to lane line detection,which greatly improves the accuracy of lane line detection.(2)A lane line detection algorithm based on self-attentive distillation and spatial convolution is proposed.SCNN(Spatial CNN)network is selected as the base network.Self Attention Distillation is combined with feedforward convolutional neural nets to enhance the learning capability of the model.This approach improves the accuracy of lane line detection for complex scenes,as well as reduces the inference time of the network to speed up the convergence of the network.Secondly,because the network has too many parameters,if all the parameters have to change once for each training,it may cause the system not to converge easily when it starts training.In this paper,a random dropout(Dropout)method is used in the fully connected layer,which can effectively avoid the overfitting caused by the system parameters updating too quickly.Finally,by comparing the experimental results obtained by adding self-attention distillation for different iteration cycles,the optimal moment to add self-attention distillation is obtained.The network has obvious advantages in convergence speed.When training the model,the training set images are intercepted at a certain ratio to increase the number of samples.The experimental results are tested with Lane Net,ENet,and SCNN networks trained on the CULane and Tu Simple datasets,respectively.The results are also compared and analyzed.The accuracy of the final algorithm on the Tu Simple dataset is96.81%,and the false detection rate and missed detection rate are 5.96% and 1.50%respectively,which is significantly better than other models.The detection speed can reach 40FPS/s when using GPU RTX2070 SUPER.The detection accuracy of the algorithm in this paper is improved on the challenging CULane dataset,which verifies the effectiveness of the proposed model in lane line detection.