Road Vanishing Point Detection by Multi-stage Convolutional Neural Network

Abstract

Road vanishing point could be used as clues to driving directions in automatic driving system. Therefore, we proposed a novel data-driven method based on convolutional neural network in order to detect the location of vanishing point of either straight or curved roads on each single given image input. Our mainly contributions are: 1. We bring advantages of deep learning technologies to vanishing point localization problem instead of using traditional manually designed texture filters; 2. a coarse-to-fine approach is proposed through cascaded heat map output CNNs. For evaluation purpose, we manually collect and annotate about 2100 road images. The experiment result on this dataset proves the robustness of our approach on almost any kinds of road type. Next we are going to integrate and test our implementation on an on-borad camera in order to perform precise navigation for robotic system such as unmanned aerial vehicle (UAV).