Research On Lane Extraction Based On Deep Learning

With the rapid development of intelligent transportation and autonomous driving,autonomous vehicles have become a global innovation hotspot and the main direction of future industry development.Its intelligent driving technology plays a key role in improving highway transportation capacity and reducing traffic accidents.The environment awareness module plays a key role in the autonomous driving technology to ensure the safety and intelligence of vehicles.As the core technology of the environment awareness system,lane line detection technology is the premise to realize the safe and autonomous driving of vehicles,and also an important means to build a high-precision map.In the actual road scene,due to the complex road conditions(such as illumination changes,diverse target shapes,and the existence of occlusion of road routes,etc.),the accuracy of lane line detection will be affected,thus affecting the safety of automatic driving.Therefore,it is also a difficulty in the current research.This paper mainly studies the detection and classification of lane lines under complex traffic scenes based on the theory of convolutional neural network.The main research contents and innovation points of this paper are as follows:In this paper,the traditional method based on Hough transform is used to detect the lane lines.The simple code compiled by C++ language is used to detect,extract and classify the lane lines.Under the simple road scene,the visualization effect of the lane line detection results is good.However,in actual complex road scenes,it is difficult to accurately detect lane lines in complex traffic scenes,such as incomplete lane lines or unpainted lane lines.Therefore,this paper proposes an end-to-end semantic segmentation network model.The proposed model,VGG-SS,is based on and optimized by VGG-16 network.By embedding self-attention distillation module between encoders and decoders and embedding SCNN module in the hidden layer at the top,the detection accuracy of the model for lane lines is improved.By training on Culane dataset,the accurate lane line detection of VGG-SS network under complex road conditions is realized.Aiming at the problems of low efficiency and low accuracy of the current traditional lane line classification method,this paper designs a simple convolutional network model FL-CNN for lane line classification.By preprocessing the Kitti data set,this paper constructs the data set for classification,and improves the training speed and generalization ability of the network through the image data set containing six different types of label data,so that the FL-CNN model can accurately classify the six different types of lane lines.At the same time,in order to obtain the distance between a vehicle and its two lane lines,this paper tests the improved Latdis Lane-1 network on on-campus self-collected data.The results show that the Latdis Lane-1 network model is improved in both accuracy and efficiency,and the final distance identification accuracy can reach the centimeter level.