Accurate lane detection with atrous convolution and spatial pyramid pooling for autonomous driving

Lane detection is a fundamental capability for autonomous driving. Many effective lane detection algorithms based on traditional computer vision and recent deep learning technologies have been proposed. However, the current state-of-the-art lane detection accuracy is still not satisfactory for realizing fully autonomous driving. Thus, this paper proposes a new lane detection network using atrous convolution and spatial pyramid pooling techniques to improve the lane detection accuracy. We address the detection problem with pixel-wise semantic segmentation. Our network consists of one encoder and two decoders, which outputs a binary segmentation map and an embedded feature map, respectively. The embedded feature map is employed for clustering algorithms to separate segmented lane pixels into different lanes. The experimental results on the public Tusimple dataset show that our network outperforms the state-of-the-arts.