Real-Time Monocular Vision-Based UAV Obstacle Detection and Collision Avoidance in GPS-Denied Outdoor Environments Using CNN MobileNet-SSD

Abstract

In this paper, we propose a monocular vision-based system that uses a MobileNet-SSD CNN for obstacle detection and collision avoidance in GPS-denied outdoor environments. This framework consists of two processes carried out simultaneously in a frame-to-frame basis: (1) an obstacle detector and classifier using a lightweight convolutional neural network with a UAV monocular onboard camera for real-time mobile systems; (2) a collision avoidance algorithm with a proportional controller responsible for the autonomous flight in GPS-denied outdoor environments. However, because object detection and classification are computationally intensive tasks, the processing is carried out off-board on a ground control station that receives online imagery and data of the UAV during the autonomous flight. The novel aspects in this work are related to the capacity of the system to detect and avoid obstacles in real-time with computationally low range hardware without GPU. We exploit public datasets meant for other purposes and carefully selected images to build a new lightweight dataset to train the CNN. Further, the output imagery data is used by a proportional controller that communicates back to the vehicle to evaluate a possible obstacle avoidance trajectory and execute it if necessary. We carried out evaluations and flights in real scenarios with multiple obstacles such as vehicles, people, bicycles, and trees for autonomous flights in GPS-denied outdoor environments with promising results.