Automatic Detection of Coastal Drift Objects Using UAV Aerial Imagery and Deep Learning for Disaster Search

Abstract

This paper describes the automatic detection of beach debris using an unmanned aerial vehicle (UAV) and deep learning to efficiently search for missing persons in large-scale disasters such as tsunamis. Using YOLO (You Only Look Once), capable of real-time detection as deep learning, we examine the detection rate of beach debris based on three years of long-term observations at a sandy beach. In addition, we describe a method that introduces background classification as a pre-processing step to apply the method to various beaches, including sand, gravel, and vegetation. The results show that an average detection rate of 85.4% can be obtained for three years between 2020 and 2022 by learning only from UAV aerial images taken two months in 2020 and that background classification can improve the detection rate by 17.2%.