一种基于FCN的车道线检测算法

A Lane Detection Algorithm Based on FCN

传统的车道线检测算法只能实现较为理想的直道情况,当应对如弯道、车道线模糊、阴影遮挡以及光线昏暗等复杂环境时,就表现得束手无策。为了解决这些问题,提出基于全卷积神经网络(Fully Convolutional Neural Network,FCN)的车道线检测算法。首先通过实车摄像头采集车道线,将车道线图像输入到全卷积神经网络进行处理,主要包括卷积、池化及反卷积等。然后经过条件随机场(Conditional Random Field,CRF)处理,得到更加精细,并具有与原始图像空间一致性的结果。最后,在大量的车道线数据集基础上,从车道线检测的质量和检测时间2方面对FCN算法的优越性进行验证。仿真结果表明,基于全卷积神经网络的车道线检测算法在检测精度、识别率和识别速度上有很明显的优越性。

The traditional lane detection algorithm can only achieve the detection of ideal straightaway,but it is helpless for complex environment,such as curves,fuzzy lane,shadows shaded on the road,dim light,etc. To solve these problems,the fully convolutional neural network（ FCN） is used. Firstly,the image of the lane acquired in real car is input into the fully convolutional neural network for processing,mainly including convolution,pooling and deconvolution,etc.In order to get more detailed result that has a good spatial coincidence with original image,the conditional random field（ CRF） is used. Finally,On the basis of the data set in the realistic scene,the advantages of FCN algorithm are evaluated from two aspects：the quality of lane detection and detection time. The experimental results show that the lane detection algorithm based on fully convolutional neural network has obvious superiority in the detection accuracy,recognition rate,and detection speed.