The “Watch” Support System for Ship Navigation

Abstract

Floating and non-floating objects such as other ships, buoys and so on must be alarmed before becoming obstacles for ship navigations. In this research, we have aimed to predict obstacles around a ship from maritime navigation images using an image recognition method and display them effectively to its operator. Faster R-CNN was used as detection method. We prepared a dataset composed of three categories for training and testing machine learning. We enumerated parameter values to obtain the best detection rate of obstacles by CNN. Then, we employed the best set of parameters for further experiments. The results are summarized as follows: (1) the detection rate of buoys is about 55 [%]; (2) large ships are sometimes mistaken for small boats. It remains to improve the detection rate and to decrease misclassifications; (3) the detection rate of small boats with distance of about 3 nautical mile(nm) from the ship is 86 [%], the detection rate of buoys with distance of about 2 [nm] from the ship is 100 [%].