Performance Evaluation of Edge Computing Object Detection Models for Maritime Surveillance on a Raspberry Pi

The exponential growth of maritime traffic has introduced significant business opportunities while also posing unprecedented challenges for surveillance. Human-based surveillance is increasingly inadequate for managing the volume and complexity of maritime activities, necessitating innovative solutions. Therefore, this study presents a comparative analysis of five prominent object detection models applied to maritime surveillance: FOMO, MobileNetV2 SSD, YOLOv5, YOLOv8, and YOLOv10. By evaluating their performance metrics in maritime environments, this research identifies the most suitable model for specific applications. For instance, YOLOv8n and YOLOv10n excel in accuracy, making them ideal when precision is paramount, irrespective of inference time and resource consumption. In contrast, FOMO is preferable when resource efficiency and shorter inference times are critical. YOLOv5n offers a balanced approach between accuracy and inference speed, suitable for scenarios requiring both performance and efficiency. This study aims to identify the most appropriate object detection model for an IoT-based maritime surveillance application focused on vessel detection, thereby enhancing operational effectiveness and decision-making. Additionally, this study proposes a novel dataset with 4,998 annotated images of 10 different ship classes, further contributing to the field.