Segmentation of Drivable Road Using Deep Fully Convolutional Residual Network with Pyramid Pooling

Abstract

In recent years, the self-driving car has rapidly been developing around the world. Based on deep learning, monocular vision-based environmental perceptions of either ADAS or self-driving cars are regarded as a feasible and sophisticated solution, in terms of achieving human-level performance at a low cost. Perceived surroundings generally include lane markings, curbs, drivable roads, intersections, obstacles, traffic signs, and landmarks used for navigation. Reliable detection or segmentation of drivable roads provides a solid foundation for obstacle detection during autonomous driving of the self-driving car. This paper proposes an RPP model for monocular vision-based road detection based on the combination of fully convolutional network, residual learning, and pyramid pooling. Specifically, the RPP is a deep fully convolutional residual neural network with pyramid pooling. In order to greatly improve prediction accuracy on the KITTI-ROAD detection task, we present a new strategy through an addition of road edge labels and an introduction of an appropriate data augmentation so as to effectively handle small training samples contained in the KITTI road detection. The experiments demonstrate that our RPP has achieved remarkable results, which ranks second in both unmarked road and marked road tasks, fifth in multiple-marked-lane task, and third in combination task. In this paper, we propose a powerful 112-layer RPP model through the incorporation of residual connections and pyramid pooling into a fully convolutional neural network framework. For small training sample problems such as the KITTI-ROAD detection, we present a new strategy through an addition of road edge labels and data augmentation. It suggests that addition of more labels and introduction of appropriate data augmentation can help deal with small training image problems. Moreover, a larger size of crops or combination with more global information also benefit improvements in road segmentation accuracy. If regardless of restricted computing and memory resources for such large-scale networks like RPP, the use of raw images instead of any crops and the selection of a large batch size are expected to further increase road detection accuracy.