一种基于高效残差分解网络的车道线检测方法

Lane line detection method based on ERFNet

为提升自动驾驶系统车道线检测的速度,提出了一种利用卷积神经网络进行特征提取,结合分类网络实现多车道线虚实线分类的方法。使用高效残差分解网络(efficient residual factorized ConvNet, ERFNet)对图像进行卷积操作和下采样,采用无瓶颈一维卷积残差结构,利用纵、横两个方向一维卷积穿插提升非线性函数的泛化性能,依据可变填充比获得多尺度上下文信息完成图像特征提取。基于反卷积与上采样结果进行特征解码,恢复原图像尺度并输出分割后的图像。相较于传统语义分割算法,本方法可减少大量特征参数,增强模型的学习能力,在提升检测速度的同时保证检测精度。在直行、转弯、上坡、下坡,道路颠簸,光照不均匀等工况下的仿真测试实验表明,本文方法检测精度可达到95.14%,检测速度较主流算法有较好提升。

In order to improve the speed of lane line detection in autonomous driving,a method of feature extraction using convolutional neural network and classification network to realize the classification of virtual and solid lane lines is proposed.An efficient residual factorized ConvNet(ERFNet)is used to perform convolution operations and down sampling on images,the network adopts a bottleneck free onedimensional convolution residual structure,utilizes vertical and horizontal one-dimensional convolution interpolation to enhance the generalization ability of nonlinear functions,obtains multi-scale contextual information based on variable fill ratios,to achieve feature extraction of images.After deconvolution and up sampling,the features are decoded and the image scale is restored,and finally the segmented image information is output.Compared to traditional semantic segmentation algorithms,this method can reduce a large number of feature parameters,enhance the learning ability of the model,and ensure detection accuracy while improving detection speed.The simulation experiments under conditions such as straight driving,turning,uphill,downhill,bumpy roads,and uneven lighting show that the detection accuracy of this method can reach 95.14%,and the detection speed is improved compared to mainstream algorithms.