Maritime Surveillance Using Instance Segmentation Techniques

A maritime surveillance system is vital in detecting objects threatening national security and violating the maritime border. Many small and non-cooperative objects, such as spyboats/refugees, take advantage of a dense marine environment to carry out illegal activities such as drug and human trafficking and being unnoticed. It broadly uses a combination of different sensor-based technologies, such as radar, visual sensors, or cameras, along with UAVs (drones) for aerial vision to accurately detect ships, but is unable to segment and obtain less detailed information. However, the computer vision community lacks suitable datasets available publicly. To address the problem, we have constructed a new ship dataset, for instance segmentation task named ShipInsSeg, which contains more than 5k marine ship/boat images that were collected and labeled manually. In this paper, instance segmentation techniques using deep learning are explored to detect and segment boats/ships with clear and precise boundaries for our dataset. We have demonstrated real-time performance and accuracy in terms of mean average precision (mAP) and frames per second (FPS).