Person classification leveraging Convolutional Neural Network for obstacle avoidance via Unmanned Aerial Vehicles

Abstract

Obstacle avoidance capability for Unmanned Aerial Vehicles (UAVs) remains an active research in order to provide a better sense-and-avoid technology. More severely, in an environment where it contains and involves humans, the capability required is of high reliability and robustness. Prior to avoiding obstacles during mission, having a high performance of obstacle detection is deemed important. We first tackled the detection problem by solving the classification task. In this work, humans were treated as a special type of obstacles in indoor environment by which they may potentially cooperate with UAVs in indoor setting. While existing works have long been focusing on using classical computer vision techniques that suffer from substantial disadvantages with respect to robustness, studies on the use of deep learning approach i.e. Convolutional Neural Network (CNN) to achieve this purpose are still scarce. Using this approach for binary person classification task has revealed improved performance of more than 99% both for True Positive Rate (TPR) and True Negative Rate (TNR), hence, is promising for realizing robust obstacle avoidance.