基于特征关联的车道线检测算法

Lane detection algorithm based on feature correlation

针对车道线检测任务中车道线细长且易被遮挡的特点,提出基于编码器解码器结构的实例分割网络——交叉卷积网络(Cross Convolution Net,C-Net),实现车道线的检测识别.首先,提出一种基于交叉卷积的特征关联机制,通过对下采样后的特征图进行连续两次的交叉卷积操作,建立单个特征点与全局特征之间的联系,增大特征图的感受野,以提高网络的推理能力.其次,采用5个双通道上采样模块对交叉卷积后的特征图进行上采样,得到车道线实例分割结果.最后,在Tusimple数据集上对网络进行训练与对比实验.研究结果表明:C-Net的准确率能够达到96.52%,且误检、漏检率较低,具有良好的车道线检测能力.

Addressing the challenges posed by the elongated and easily occluded nature of lane lines in lane de-tection tasks,this study introduces the Cross Convolution Net(C-Net),an instance segmentation network structured on an encoder-decoder architecture,for effective lane detection and recognition.Firstly,a feature as-sociation mechanism based on cross convolution is proposed.Through two consecutive cross-convolution op-erations on the down-sampled feature map,a connection is established between individual feature points and the global features,thereby enlarging the receptive field of the feature map to enhance the network’s inferential capabilities.Furthermore,5 dual-channel up-sampling modules are used to up-sample the cross-convolution feature map,yielding the instance segmentation result of lane lines.Finally,the network is trained and com-pared on the Tusimple dataset.The results show that C-Net can achieve an accuracy rate of 96.52%,with low false detection and missed detection rates,highlighting its robust lane detection capabilities.