見張りの自動化に向けた深層学習での領域分割と遠方船舶の認識

抄録

  One of the technologies required for fully autonomous ship operations is an unmanned/automated watch-keeping system using camera images. Methods to detect objects, such as ships, from images have long been studied. However, it has been difficult to recognize the target when it is surrounded by complex background or it appears tiny due to long distance. In addition, the environment surrounding the ship, such as the land, harbors and quays, also needs to be recognized appropriately. This paper proposes a method to recognize the peripheral area with camera images utilizing deep learning. It then describes the results of three experiments we conducted by dividing the surrounding area into the sea, land, sky and ship segments. Using these segmentation results, we partially enlarged the images to reexamine the existence of ships in the background. The experimental results successfully identified more ships that appeared a long distance away after enlarging the images of the border areas between sea and sky/land.