Cross Talk 3 < FURUNO ELECTRIC CO., LTD.> Challenge to the practical use of automatic detection by AI image analysis
Following in the footsteps of the automobile industry, where automated driving is becoming more and more common, the world of shipping has also begun to work towards automated navigation. As part of the unmanned navigation demonstration to achieve this goal, pluszero has been working with Furuno Electric, a global manufacturer of marine radars, on research and development of AI-based image analysis.
Cross Talk 3＜古野電気＞ トレーラー
Ships are entering an era of unmanned operations. What are the challenges that stand in the way of realizing this?
Furuno Electric is a global manufacturer of marine radar-related equipment, and is one of the companies with the largest share in this field in the world. Furuno Electric and pluszero are currently working together on a demonstration experiment for unmanned navigation. Please tell us why you chose us as your development partner.
In the world of automobile industry, we all know that "automatic driving systems" are almost ready for practical use. In the same way, in the world of ships, efforts to build unmanned ships that can navigate safely have become active in recent years. The demonstration experiment that we are currently working on was solicited by the Nippon Foundation, which aims to commercialize unmanned ships by 2025, and we have decided to participate as a representative of a marine electronics manufacturer.
A ship's radar is a very important safety device that displays images of obstacles around, where other ships are, and whether there is a risk of collision. In order to make it unmanned, a very high level of image analysis technology using AI is required. A user of our equipment and who we were trusting very much introduced us to pluszero. The fact that pluszero is an "up-and-coming AI venture company with engineers by the University of Tokyo" also convinced us that they would be able to help us solve the difficult problems we were facing in this project.
The problem that your company was facing was the appearance of "false images" in the radar images, images that do not actually exist. This was a challenging and rewarding task for us, as we had to replace the radar images and binoculars with AI that would normally be used by navigators to make visual judgments.
Yes, that's right. The radar fires its own radio waves and captures the reflected waves to display images of ships and other obstacles around the ship. However, there are times when a "reflector" appears on the radar screen, reflecting the radio waves like a mirror, even though the ship does not actually exist.
Skilled navigators know from experience and knowledge of geography, weather, and other conditions that make it easy for "false images" to appear, so they can immediately determine whether this is a real ship or a false image. On the other hand, it is very difficult for inexperienced navigators to identify false images, and AI-based identification of false images is highly sought after by such users.
In this project, we had to start from building the data itself, rather than having all the necessary data from the beginning and just letting AI learn from it. One of the important points in creating data is to carefully consider how the AI will learn the data. The difficulty of using AI lies in creating appropriate data with this in mind. In this case, we worked together from the base data construction while maximizing the combination of Furuno Electric's radar knowledge and our AI knowledge.
Detecting "False Images" on Radar Screens! What are the technologies and innovations that lead to the solution of the problem?
In this project, we actually installed radar on a ship operated by a major shipping company to collect data in various situations, and the data accumulation has been going relatively smoothly. For the subsequent reconstruction of the data and learning by AI, pluszero proactively made various technical proposals and innovations.
We have learned that there are various types and shapes of false images, and that some are easy for AI to identify and others are not. We need to be able to detect the ones that are difficult to distinguish at a practical level, while letting the AI learn appropriately that these are false images.
This is where we use a technology called "semantic segmentation". In brief, it is a technique to appropriately cut out the objects that we want to detect in an image, and it is already widely used in the world. The important thing to remember is that some false images cannot be determined from just one image. In order to detect this type of false image, it is necessary to look at multiple images in a time series and make a comprehensive judgment. For this reason, semantic segmentation is performed after incorporating time-series information as well.
As you say, when navigators actually judge false images on the radar screen, they are not looking at a single picture but rather the movement of images in a time series. If we can incorporate the knowledge and know-how of how humans determine false images into AI, I have high hopes that we can eventually achieve the same or even better performance than that of skilled navigators.
Yes, in order to make good use of AI, it is necessary to have the knowledge to select the appropriate method from a lot of options according to the problem. I think it is very important to have a relationship with people who can think about the next step together while sharing information about what works and what doesn't, what causes it, and what kind of innovations are possible. I think this project with Furuno Electric is an example of how we can understand each other and work together very well.
In this project, Furuno Electric, a marine radar company, and pluszero, an AI company, were able to conduct research and development with specialists from unexpectedly different fields. I would like to continue to make great achievements toward our goal of achieving unmanned navigation.