Research On Lane Detection Algorithms Based On Deep Learning

Lane detection is the identification and analysis of lane on the road by computer vision technology.Lane detection provides important road information for autonomous driving systems and provides a reliable basis for autonomous driving to achieve high precision driving decisions,and the task has been an important part of visual perception for autonomous driving.With the development of autonomous driving technology,lane detection as a research hotspot and is widely used in advanced driver assistance systems(ADAS)such as assisted road changing and intelligent obstacle avoidance.This research is based on deep learning on lane detection as the research object and addresses the problem of accuracy degradation caused by environmental factors such as darkness and shadows in lane detection algorithm and solves the problem of robustness standardization of the test model by designing a series of robustness testing methods and indicators during the experiment.This paper analyzes and tackles some of the difficulties existing in lane detection and possible problems in the real environment based on current deep learning and shows the effectiveness of the methods and experiments proposed in this topic in several comparative experiments through theoretical research and experimental validation,and its main work and innovations are as follows.First,the road environment and lighting conditions often interfere with the lane detection algorithm,making the algorithm’s detection in complex road situations unsatisfactory.To improve the algorithm’s detection in these situations,this paper combines the powerful feature extraction capability of convolutional neural networks and the multiplicative attention mechanism in Transformer architecture and applies the self-attention structure to convolutional neural networks,which improves the detection effect of deep convolutional neural networks in situations such as darkness by making the model pay more attention to the important parts of the road.Second,to better compare the robustness differences between different lane detection algorithm models in deep learning neural networks.This paper detects the robustness of the models under different situations by designing a series of interference experiments using black-box testing based on various unexpected situations and disturbances that may occur in the real environment,both in terms of signal collection and physical environment.In addition,a new robustness metric is proposed in this paper to better evaluate the robustness of the model.Third,deep neural networks inference speed problem are usually encountered when deploying network models on resource-limited devices.Aiming to achieve a lightweight model structure by compressing the model using methods such as knowledge distillation has received a lot of attention from academia and industry.To better meet the real-time performance of lane detection,this paper designs a knowledge distillation method with effective information,and attempts to achieve model light weighting through knowledge distillation,so that the accuracy of the light weighted model will not be significantly degraded,thus improving the effectiveness of model deployment in realistic environments.