Research On Lane Detection And Classification Algorithm Based On Convolutional Neural Network

In recent years,due to the increase in the number of cars,the incidence of traffic aceidents has gradually inereased,and the:issue of automobile driving safety has beeome a hot issue in society.Driverless cars can effectively reduce the occurrence of traffic accidents,so more and more researchers are paying attention.Lane detection and classification is an important part of driverless cars.Detecting lane quickly and accurately is the goal pursued by researchers.Traditional lane detection methods are mainly based on features or models.When faced with lane damage,illumination,and road slope changes,the algorithm's robustness decreases,its adaptability is poor,and its detection accuracy decreases accordingly.The success of deep learning methods in the field of computer vision provides new ideas for lanes detection and classification.By constructing a neural network model to automatically learn picture features,not only the detection speed and aceuracy are improved,but also the generalization ability of the algorithm is improved.This paper researches lanes detection and classification based on convolutional neural network.The main work is as follows.Firstly,in the process of lanes extraction,aiming at the problem of low utilization of multi-dimensional image information of an efficient semantic segmentation neural network(ENet),resulting in low accuracy of lane segmentation,this paper improves ENet and proposes efficient residual neural network(ERNet).Two processing streams are used for feature propagation:one processing stream is a feature extraction stream used to obtain high-dimensional semantic information,which we call a pooled stream;the other processing stream is a residual stream used to record low-dimensional boundaries information.The Lurpose of introducing the residual strearn is to strengthen the propagation of low-dimensional boundary features and encourage feature reuse of low-dimensional boundary features.Experimental results show that on the CamVid dataset,Cityscape dataset,SUN RGB-D dataset and Tusimple dataset,the segmentation speed of ERNet is close to that of ENet,but the segmentation accuracy exceeds ENet.Secondly,in order to solve the problem that the traditional fitting method uses the same perspective transformation matrix as the flat road when the vehicle is going up and down,which causes the fitting accuracy to decrease and even the fitting deviation to occur,this paper proposes a method of lane fitting based on a eonvolutional neural network.Construct a convolutional neural network for training,make the network output variable perspective transformation parameters,combine maps of lane segmentation,perform perspective transformation on the lane,and then use the least square method to fit the lane,and finally obtain the visible graph of lane fitting based on the original image.Experimental results show that the method in this paper can effectively improve the effect of lane fitting.Finally,because the traditional lane classification method can only classify lanes with different virtual and solid lanes,it cannot classify lanes with different colors and virtual and solid lines,with slow classification speed.Aiming at this problem,t,his paper proposes a lane classification method based on a convolutional neural network,which is trained based on the VGG11 model,and the network is compressed at the same time.The model volume is compressed while ensuring the classification accuracy,and the classification speed is improved.The experimental results prove that the compression model based on VGGll has been compressed nearly 6 times,and the running speed has been increased accordingly.The final output is five lane categories with different colors and virtual and solid lines.