An Explainable CNN-based Approach for Maritime Search and Rescue On Edge

Saving lives at sea remains central to maritime search and rescue (SAR) missions. Traditional methods such as aerial and marine visual searches, helicopter, radar and sonar systems are inefficient, costly, and less effective when dealing with small or hard-to-detect objects. Unmanned aerial vehicles (UAVs) have emerged as a powerful tool to improve response times to save more lives. In this work, a custom convolutional neural network (CNN) was developed and trained on the SeaDronesSee dataset to detect stranded people or boats in UAV-captured video over the sea. Our model obtained an accuracy of 68.4 percent the challenging SeaDronesee dataset with real-time performance required for low-powered computers such as Jetson Orin. When deployed on the Jetson AGX Orin platform operating at 50W, the model achieved a speed of 32 frames per second.