Maritime UAV Patrol System Utilizing YOLO Framework and Darknet-53 for Object Detection and Alarm System

Maritime patrol personnel usually perform patrol tasks via ships and helicopters in the search and rescue work at sea, which often consumes many resources with human and large patrol tools. The victims call for help, and the rescue unit obtains the ship's information but lacks maritime images of the sea and thus cannot control the accident site in time. In this article, we propose an automatic patrol task based on small UAV and different maritime search patterns corresponding to different sea areas. The search patterns include parallel and expanded squares to be applied on the coast and sea and used to assist in searching for victims. Maritime images are collected via a deep neural network-based UAV and returned to the search and rescue system in time. The object detection model is trained using images of maritime accidents collected from the network. Maritime accidents can thus be detected by the trained object detection model based on the YOLO framework. The YOLO uses the darknet53 model as the backbone to have the advantage of quickly and accurately identifying objects. The severity of a maritime accident is defined according to the risk assessment value of the event. We divide the severity of the event into two indicators: rescue and monitoring, called the danger level. In our work, maritime accidents are divided into two classes: humans in the sea and ship collisions. The risk analysis of maritime accidents is used to distinguish whether the objects need to be rescued or only monitored. We setup an alarm system to classify the danger level of the targets and perform rescues according to the corresponding levels. Finally, the Short Message Service (SMS) sends an alarm message to the rescuers. In our experiments, we set 100 epochs to train the model and compare different hyperparameters: iteration, learning rate, batch size, and image size. The best results of the iterative learning rate, batch size, and image size are 2000, 0.001, 64, and 608, respectively. The proposed system can detect humans in the sea and ship collisions with an accuracy of 94.20% (mAP), which is 11% more accurate than the Single Shot MultiBox Detector 512 (SSD512) method used for object detection, and immediately classify the danger level of the accident to notify the rescuers.