A Deep Convolutional Networks for Monocular Road Segmentation

Abstract

free space detection is a very important task in the autopilot system. In this paper, for traditional monocular RGB images, we propose an accurate road detection system based on deep convolutional networks (CNN). Compared with classification, semantic segmentation needs to predict each pixel, so it is computation expensive and hard to achieve realtime. In our road detection system, we reduce the forward time when the detection accuracy is close to other state-of-the-art algorithms. These benefits from the faster forward network we use and the Multi Detection Model we designed. Modern deep convolutional network architecture becomes wider and deeper because they can improve network performance. According to this idea, we proposed a Multi Detection Model (MDM) for applying in our road detection system. With the input of high resolution (375×1242), our network is trained end-to-end. Compared with the other state-of-the-art networks in the KITTI road detection, we shortened the calculation time when the detection accuracy was close to.