Research On Lane Detection Technology Based On Convolutional Neural Network

Lane detection technology is a key technology in autonomous driving.It provides an important basis for the car's lane keeping,direction guidance and other functions.With the development of technologies related to autonomous driving,higher requirements are placed on lane detection technology.The detection of lanes not only needs to adapt to changing and complex real scenes,but also needs to achieve real-time and accurate detection around the clock.The traditional lane detection method is difficult to meet the actual application needs,so in recent years,the lane detection technology based on deep learning has become a research hotspot.Using convolutional neural network to detect lanes can effectively improve the algorithm's ability to respond to changes in road scenes.Semantic segmentation is a commonly used lane detection method in the field of deep learning.However,current relevant research has problems such as low utilization of semantic information and vague feature expression.Therefore,based on the frontier work,this paper continues to study the lane detection technology based on semantic segmentation.The main work is as follows:(1)A lane detection method based on semantic segmentation and pixel feature instance is proposed.In order to achieve real-time and accurate detection,a multi-task lane detection model of codec architecture is designed based on the improved ERFNet as the basic network.In this model,effective road segmentation,lane segmentation and pixel feature discrimination tasks of each lane are jointly trained to obtain the instance results of lanes in the original figure.In the post-processing stage,morphological processing is used to optimize the semantic segmentation results,and clustering algorithms are used to analyze lanes.Experiments show that the accuracy of this method is higher than that of Lane Net,SCNN and other methods on public dataset.All indicators are better than Lane Net on in-house dataset,and the detection speed reaches20 fps,which meets the real-time requirements.(2)A lane detection method based on semantic segmentation and spatio-temporal correlation information is proposed.Existing detection methods mostly use single-frame images,while lanes are continuous linear structures,and adjacent frames contain rich relevant information.Therefore,our paper proposes to combine the spatio-temporal correlation information of consecutive frames to optimize the lane segmentationbranch.We embed a long-short-term memory network(Long Short-Term Memory,LSTM)module in the model to build an end-to-end network.This module uses LSTM to fuse the information of previous multi-frame images to predict the lanes in the current frame.The experimental results show that the optimization of the lane segmentation branch proposed in this paper is indeed effective.The effect of this method on the public dataset(Tusimple)is relatively improved.In addition,when verified on a challenging dataset,it is found that the improved algorithm has better accuracy and robustness than the original method.In this paper,a large number of experiments are carried out on public dataset and in-house dataset,which verify the effectiveness and practicability of our lane detection method based on convolutional neural network.