Automatic Training Data Generation Method for Pixel-Level Road Lane Segmentation

摘要

Lane detection or road detection is one of the key features of autonomous driving. By using deep convolutional neural network based semantic segmentation, we can build models with high accuracy and robustness. However, training a pixel-level semantic segmentation needs very tine-labeled training data, which requires large amount of labor. In this paper, we propose an automatic training data generating method, which can significantly reduce the effort of the training phase. Experiments prove that our method can generate high-quality training data for lane segmentation task.