Autonomous Outdoor Building Navigation Using a Single Monocular Camera

In this work, we describe an autonomous robot system that navigates an outdoor building environment using color monocular images from a single camera. This system is able to avoid dynamic obstacles, such as pedestrians, and recognize its location in terms of which hallway it is located in. Using a multi-task convolutional neural network, the system processes images in real-time and produces predictions for 4 tasks: topological robot localization, driveable space classification, intersection detection, and hallway goal prediction. These predictions allow the robot to determine if an area is free of obstacles and allows the system to plan a safe, driveable path. We outline how training data is collected for each of the tasks, describe the overall neural network architecture, and cover what each network output head produces. We find the system can robustly traverse a limited outdoor building scenario at various times of day and lighting conditions.