基于深度学习的低光环境车道线检测算法仿真  

Simulation of Lane Detection Algorithm in Low-Light Environment Based on Deep Learning

作　　者：张琰[1,2] 赵庆 梁莉娟[1] ZHANG Yan;ZHAO Qing;LIANG Li-juan(Information Engineering Institute,Wuchang Institute of Technology,Wuhan Hubei 430065,China;Universitéde Technologie de Compiègne,Dans le Department de L′Oise en Région Picardie Compiègne 60200,France)

机构地区：[1]武昌工学院信息工程学院,湖北武汉430065 [2]法国贡比涅技术大学(UTC),皮卡第大区瓦兹省贡比涅60200

出　　处：《计算机仿真》2024年第5期152-157,共6页Computer Simulation

基　　金：湖北省高等学校优秀中青年科技创新团队计划项目(T2021042);武昌工学院校级学科群“新能源与智能物联”(2021XK01)。

摘　　要：车道线检测研究是保证车辆自动驾驶安全的基础,但当下研究中存在低光环境车道线检测稳定性差,准确率低的问题,为此提出一种基于改进BFA-Retinex光照补偿深度学习算法,提高图像重点区域的光照水平,通过提取并融合车道线纹理与直方图特征,构建出BRI-SVM低光车道线检测识别模型。模型包括图像补光模块、特征融合模块与车道线检测模块三个模块。BRI-SVM模型首先对Lll低光车道数据集进行灰度化与标准化处理,然后采用改进双边滤波算法提高图像光照基准;接着提取优化图像中的H与G特征,并将二者有机融合;最后基于数据驱动的方法,以深度学习与SVM算法为核心,构建出BRI-SVM模型,并采用交叉验证的方式提升模型性能。多类融合算法模型的仿真结果表明,在Lll低光数据集上,与其它模型相比,BRI-SVM模型的稳定性能与综合性能最高,特征值分别达到96.1%与96.3%,较传统算法分别平均提升了24.4%和23.8%;此外,所构建的模型具有较好的检测时效性与检测准确性,在所有模型评价中排名第2。综上所述,基于改进BFA-Retinex算法的低光照环境下车道线检测模型在具有最高鲁棒性与稳定性的同时,大幅度提高了车道线检测的准确性与时效性。Lane line detection research is the basis to ensure the safety of vehicle automatic driving,but there are some problems in the current research,such as poor stability and low accuracy of lane line detection in low-light environment,so this paper proposes a deep learning algoritm based on improved BFA-Retinex illumination compensation to improve the illumination level of key areas of the image.The BRI-SVM low-light lane line detection and recognition model was constructed,which includes three modules:image light filling module,feature fusion module and lane line detection module.Firstly,the Lll low-light lane data set was grayed and standardized by the BRI-SVM model,and then the improved bilateral filtering algorithm was used to improve the image illumination benchmark,and then the H and G features in the optimized image were extracted and organically fused;Finally,based on the data-driven method,the BRI-SVM model was constructed with the deep learning and SVM algorithm as the core,and the performance of the model was improved by cross-validation.The simulation results of the multi-class fusion algorithm model show that,on the Lll low-light data set,compared with other models,the BRI-SVM model has the highest stability and comprehensive performance,with the eigenvalues reaching 96.1%and 96.3%,respectively,and an average increase of 24.4%and 23.8%,respectively,compared with the traditional algorithm;In addition,the model constructed in this paper has good detection timeliness and detection accuracy,ranking second in all model evaluations.To sum up,the lane detection model based on the improved BFA-Retinex algorithm in low-light environment has the highest robustness and stability,and greatly improves the accuracy and timeliness of lane detection.

关 键 词：光照补偿 特征融合 车道线检测 

分 类 号：TP391.9[自动化与计算机技术—计算机应用技术]