Author(s)

Xiongxuan Huang and Xueyun Chen

Corresponding Author

Xiongxuan Huang

ABSTRACT

Deep learning methods have achieved excellent detection results in various road databases. However, due to some harsh road conditions, these network models still suffer from some abnormal phenomena such as hollowness, unsmooth edges and so on. To address this problem, we improve the road detection network via a novel convolutional-kernel-based conditional random field (CK-CRF), which designs some special convolutional kernels for hollowness scanning. Experiments on two open road datasets show that the proposed method outperforms the-state-of-the-art models by an obvious margin, including the famous deeplab network with a traditional conditional random fields based on fully-connected layers.

KEYWORDS

Automatic pilot; deep learning; conditional random field; convolutional neural network