A Deep Learning Method to Obstacle Avoidance for Disaster Investigation Robot

Abstract

The objective of the research is to realize a disaster investigation robot that can avoid obstacles autonomously only by using a single camera. Obstacle avoidance system is divided into two parts. One is a trained Deep Fully Convolutional Network for making a depth map from a single image without a need to compute a global map. The other is a height and distance estimation of obstacles for climbing up by the robot, in which the following methods are used: a Convolutional Neural Network for obstacle detection, a height calculation using a bounding box and a distance estimation by a focal length of the camera. As a goal, hardware experiments will be achieved by implementing the methods into the real robot.