Detecting beach litter in drone images using deep learning

Abstract

Beach pollution through litter leads to various negative effects and mitigation and cleanup efforts are required to prevent accumulation. Automated monitoring is an important step towards this goal and has become much more feasible with recent developments in drone technology and artificial intelligence / object detection. To assess the potential of artificial intelligence for litter monitoring on Maltese beaches, two deep learning algorithms (YOLOv5 and Faster R-CNN) were trained on drone footage of beach litter and their detection performance compared. With a mean average precision (mAP) of 0.542 YOLOv5 outperformed Faster R-CNN (mAP: 0.328). In addition, detected litter objects were geolocated so that their position could be estimated with an average error of 3.7 meters.