摘要
基于传统图像处理方法的车道线检测算法,易受到各种自然条件的影响,鲁棒性差、检测准确率不高,且不具备特征的语义描述能力,当图像的像素发生变化时检测效果会明显下降。针对这些问题,为提高复杂场景的车道线检测准确率,提出一种基于多尺度全卷积神经网络的车道线检测模型。该模型首先采用空间金字塔池化结构提取多尺度的图像纹理信息来增强深度网络的表征能力;其次选用加权损失函数提高车道线的检测准确率;最后通过训练网络选取最优参数和方法来使其最优,完成端到端的道路场景语义分割,以实现车道线的检测。在Tusimple数据集上进行测试,结果表明该模型具有较好的车道线检测能力,在多场景条件下车道线检测准确率能达到95.56%,能有效辅助汽车驾驶。
The lane detection algorithm based on traditional image processing method is susceptible to various natural conditions,which may cause poor robustness and low accuracy of detection.Additionally,it is not able to semantically describe the detected features.When the pixels of the image change,the detection quality will decline obviously.In order to solve these problems,a lane detection model based on multi-scale full convolutional neural network was proposed.Firstly,spatial pyramid pooling structure is used to extract multi-scale image texture information to enhance the representation ability of depth network;secondly,weighted loss function is used to improve the accuracy of lane detection;finally,optimised parameters and methods are selected by training network to achieve end-to-end road scene semantics segmentation,to achieve lane detection.The results of experiments on Tusimple dataset reflect that the model has a good performance on lane detection.The accuracy of lane detection in multi-scene can reach 95.56%.
作者
王帅帅
刘建国
纪郭
WANG Shuaishuai;LIU Jianguo;JI Guo(Hubei Province Key Laboratory of Modern Automotive Technology,Wuhan University of Technology,Wuhan 430070,China;Hubei Collaborative Innovation Center for Automotive Components Technology,Wuhan University of Technology,Wuhan 430070,China)
关键词
全卷积神经网络
车道线检测
语义分割
多尺度
空间金字塔池化
full convolution neural network
lane detection
semantic segmentation
multi-scale
spatial pyramid pooling
作者简介
王帅帅(1993-),男,江苏苏州人,武汉理工大学汽车工程学院硕士研究生.
