A Convolutional Neural Network Vision System Approach to Indoor Autonomous Quadrotor Navigation

Abstract

A Convolutional Neural Network (CNN) vision-based approach is demonstrated to enable autonomous flight of a stock unmodified quadrotor drone in hallway environments. The video stream from a monocular front-facing camera on-board a quadrotor drone is fed to Convolutional Neural Network (CNN) environment classifiers at a base station in order to detect upcoming intersections and dead-ends. Detecting these hallway structural features allows our control planning algorithms to take appropriate action in order to stop and turn at intersections or stop before colliding with dead-ends such as walls and doors. The use of CNNs permit intersections and dead-ends to be detected with a high degree of accuracy in a wide variety of indoor environments with varying contrasts, lighting conditions, obstructions, and many other conditions that prevent easy generalization of feature extraction. Overall, our approach allows for real-time navigation at high rates of speed approaching 2 m/s.