Automatic obstacle avoidance of quadrotor UAV via CNN-based learning

Abstract

In this paper, a CNN-based learning scheme is proposed to enable a quadrotor unmanned aerial vehicle (UAV) to avoid obstacles automatically in unknown and unstructured environments. In order to reduce the decision delay and to improve the robustness for the UAV, a two-stage end-to-end obstacle avoidance architecture is designed, where a forward-facing monocular camera is used only. In the first stage, a convolutional neural network (CNN)-based model is adopted as the prediction mechanism. Utilizing three effective operations, namely depthwise convolution, group convolution and channel split, the model predicts the steering angle and the collision probability simultaneously. In the second stage, the control mechanism maps the steering angle to an instruction that changes the yaw angle of the UAV. Consequently, when the UAV encounters an obstacle, it can avoid collision by steering automatically. Meanwhile, the collision probability is mapped as a forward speed to maintain the flight or stop going forward. The presented automatic obstacle avoidance scheme of quadrotor UAV is verified by several indoor/outdoor tests, where the feasibility and efficacy have been demonstrated clearly. The novelties of the method lie in its low sensor requirement, light-weight network structure, strong learning ability and environmental adaptability.

A Survey of Computer Vision Methods for 2D Object Detection from Unmanned Aerial Vehicles

“…The computational burden of these models is still high when applied to the UAV’s obstacle avoidance. Also, it is proved that part of the computation is not required since the blocks with a smaller number of layers could yield reliable results for the UAV’s obstacle avoidance [ 101 ].…”

“…Moreover, authors report how activation maps show the focus of the network on objects such as street lines that are an indicator of steering directions. Differently from deep learning-based control approaches that focus on reinforcement learning schemes, this family of works relies still on supervised learning, and it represents a very active research field [ 101 ].…”

Towards Fully Autonomous UAVs: A Survey

“…Additionally, many of the existing approaches consider only generating motion decisions/policies in 2D without utilizing the full maneuverability of UAVs. Most of these methods are based on deep reinforcement learning [ 136 , 137 , 138 , 139 , 140 , 141 , 142 , 143 ] and deep neural networks [ 144 , 145 , 146 , 147 , 148 , 149 , 150 , 151 ].…”

Control Strategies and Novel Techniques for Autonomous Rotorcraft Unmanned Aerial Vehicles: A Review

“…Various research works have been conducted to enhance the collision avoidance control and efficiently improve responses for any known or unknown obstacles. Dai et al [188] introduced an automatic obstacle avoidance system based on a convolutional neural network (CNN) for a quadcopter system to automatically avoid obstacles and fly safely and efficiently in unknown indoor/outdoor environments. The method revealed several advantages including low sensor requirements, strong learning capability, light-weight network structure, and environmental adaptability.…”