Automatic detection of rescue targets in maritime search and rescue missions using UAVs

Unmanned Aerial Vehicles (UAVs) can be an important resource when performing Search and Rescue (SAR) operations at sea, as this technology is fairly inexpensive when compared to traditional SAR approaches that use significant human resources and expensive air and naval assets, thus enabling the deployment of several UAVs simultaneously in these missions to perform rescue targets search in maritime environments.In order to maximize the usefulness of these UAVs in such operations, we propose a method which utilizes a state-of-the-art object detection network to perform real-time rescue target detection on-board the UAV, using standard RGB cameras, with minimal human intervention, thus enabling an increased vehicle autonomy and search range. Additionally, since the UAVs only relay the candidate images and locations that contain possible rescue targets, given by the onboard detector, it is possible to have several UAVs working in parallel that report back to a single human operator.We have selected the YOLOv4-tiny detection network, pretrained in the COCO dataset, and retrained it to detect rescue targets at sea. For this purpose some datasets were recorded and annotated to simulate the presence of maritime rescue targets. The proposed approach has been validated on an independent test dataset, showing that it has good detection capabilities and thus providing convincing results regarding the use of UAVs with automatic target detection capabilities in SAR missions.