首页|基于生成式对抗网络与ResNeXt的车道线检测算法

基于生成式对抗网络与ResNeXt的车道线检测算法

扫码查看
现实场景中汽车行驶环境具有复杂多变性,光照变化、道路阴影、车辆及建筑物遮挡会对车道线的识别造成干扰.针对该问题提出一种生成式对抗网络与深度神经网络架构(ResNeXt)相结合的语义分割车道线检测算法.首先采用高斯滤波器和线性点运算结合的方法对输入图像进行预处理,增强图像的纹理特征,降低光照及噪声对图像的影响.其次采用生成对抗网络SAGAN,生成图像来扩充数据集,并结合ResNeXt网络构成SGRNeXt检测模型,该模型采用VGG堆叠的思想和Inception的Split-Transform-Eerge思想,在不增加参数复杂度的前提下提高准确率,减少超参数数量,同时采用基于行方向上的位置选择、分类的算法,在全连接层上进行分类,并使用全局特征作为提取特征来解决感受野的问题.而后进行语义分割的车道线检测,最后在图森数据集上进行测试和验证.试验结果表明,本研究所提出的SGRNeXt检测算法的准确率可达 95.7%,基于行方向上的位置选取与分类算法使该模型在识别速度上具有明显的提升,FPS可达 53.74,满足实时性要求.生成式对抗网络的引入,可以使模型更加稳定地训练,防止过拟合化,增强了模型分类能力.本方法在具有视觉遮挡和多变照明条件下对车道线的识别具有很好的检测效果,提升了车道线识别在多样环境下的鲁棒性和准确性.
Traffic Line Detection Algorithm Based on Generative Adversarial Networks and ResNeXt
In the real scene,the driving environment is complex and changeable.The illumination variation,road shadow,and occlusion of vehicles and buildings interfere with the traffic line identification.Due to this issue,the semantic segmentation traffic lane detection algorithm based on the combination of SAGAN and ResNeXt was proposed.First,Gaussian filter and linear point operation were used to preprocess the input images.The image texture features were enhanced.The influences of illumination and noise on the images were reduced.Second,the generation countermeasure network SAGAN was used to generate the images for expanding the data set.Combined with ResNeXt network,the SGRNeXt detection model was established.The model adopted the idea of VGG stacking and the Split-Transform-Eerge idea of Inception,which could improve the accuracy and reduce the hyper parameters amount without increasing parameters complexity.Simultaneously,the algorithm,based on row-direction position selection and classification,was used to classify on the full connection layer.It used the global features as extraction features to solve the problem of receptive field.Then the traffic line detection based on semantic segmentation was carried out.Finally,the test and verification were performed on the TuSimple data set.The result indicates that the accuracy of proposed SGRNeXt detection algorithm can reach 95.7%.The position selection and classification algorithm,based on the row direction,makes the recognition speed significantly improved,and the FPS can reach 53.74,which meets the real-time requirements.The addition of SAGAN can make the model more stable,prevent overfitting,and enhance the model classification ability.The proposed algorithm has the good detection effect on traffic line recognition under visual occlusion and variable lighting conditions.It can improve the robustness and accuracy of traffic line recognition in various environments.

intelligent transporttraffic line detectionSGRNeXt modeldeep learningautomatic drivingResNeXt

潘玉恒、刘泽帅、鲁维佳、汪佳、李慧洁

展开 >

天津城建大学 计算机与信息工程学院,天津 300384

智能交通 车道线检测 SGRNeXt模型 深度学习 自动驾驶 ResNeXt

2024

公路交通科技
交通运输部公路科学研究院

公路交通科技

CSTPCD北大核心
影响因子:1.007
ISSN:1002-0268
年,卷(期):2024.41(12)