Obstacle Detection for Rover Navigation Using Semantic Segmentation

This paper considers the problem of navigation of an autonomous mobile robot (rover) across unfamiliar, rough terrain using modern semantic segmentation methods. More specifically, the problem of searching for and recognizing obstacles that arise along the rover’s path is investigated. A review and comparison of modern neural network architectures designed for semantic segmentation, such as fully convolutional network, SegNet, and U-Net, are presented in relation to the stated problem. Computational experiments were performed on the Robot Unstructured Ground Driving Dataset with a modified number of classes. Using the U-Net architecture, a segmentation accuracy of 94.84% was achieved on this dataset according to the Intersection over Union metric.