Abstract:
Maritime Domain Awareness (MDA) is essential for many applications in both civil and
military fields in order to avoid non-traditional threats such as drug trafficking, illegal
unreported and unregulated fishing and human smuggling. Real time ship detection mechanisms
are needed for MDA and conventional ship detection mechanisms are not efficient, reliable, and
economical due to the complexities in the sea and large scale human interventions. Therefore,
there is a high demand for automated on-board ship detection mechanisms that can operate on
a ship, unmanned aerial vehicle (UAV) or shore. The overall objective of this work is to lay the
stepping stones to build an automated real-time ship detection system for MDA. Here, popular
MobileNet V2 and MobileNet V2 - FPN lite machine learning models are trained, deployed and
tested. They have the portability to run on portable embedded systems smoothly. In this work,
the training data set is prepared using publicly available Electro Optical (EO) dataset and Near
Infra-Red (NIR) dataset. Both models are deployed in a portable embedded platform called
Jetson Nano. In our experiments, MobileNet V2 and MobileNet V2 - FPN lite models show
mean Average precision (mAp) of 0.708 and 0.882. The Frames Per Second (FPS) rates of the
video streams in MobileNet V2 and MobileNet V2 - FPN lite are 40 FPS and 5 FPS.