Computer vision is already guiding cranes – Research aims for increasingly accurate navigation

The University of Helsinki starts to develop a positioning solution based on machine learning together with the crane and technology company Konecranes.

Researchers from the University of Helsinki are launching a project to develop a positioning system based on machine learning and an increasingly accurate and inexpensive navigation solution in cooperation with Konecranes. The company donated €160,000 to the University for the project.

The project is to be launched in early 2020. The University of Helsinki and Konecranes will be developing the capacity of the cameras installed in lifting equipment to observe their surroundings and identify various objects automatically. This requires methods associated with both computer vision and deep learning.

The project focuses on a computer vision technique known as visual SLAM (simultaneous localisation and mapping). For instance, a machine moving about a construction site will be constantly aware of its position with the help of cameras utilising this technology. The cameras will also generate a continuously up-to-date map of the surroundings.

Human movement is difficult to anticipate

Construction sites are good locations for developing navigation technology, as the surroundings are in constant flux and there are a large number of moving objects. Certain surprising elements, such as moving human beings, bring an additional challenge to the mix.

“Machines must know where to go at the construction site. They also need to be able to take a hold of objects with precision and lift them up safely. At the same time, the machines have to observe their surroundings to avoid accidents,” says Laura Ruotsalainen, associate professor of computer science and head of the project.

Ruotsalainen has previously investigated indoor navigation and ways to make it more precise through computer vision. She has been looking into how an accurate view of the surroundings can be gained even from low-quality positioning data collected with a number of devices by compensating for the quality with computer vision techniques.

“In this study, we will have the chance to work in a research laboratory simulating a genuine factory environment. First, we will collect data and identify the biggest problems. After that, we will start developing the visual SLAM technique and deep learning to tackle these problems,” Ruotsalainen says.

“We are glad to embark on this collaboration with the University of Helsinki. It will support many goals central to our business, such as investing in technological solutions to ensure our pioneering status on the market. We also wish to develop increasingly safe work environments and strengthen our ties to the educational institutions that are educating our future employees,” states Matti Kemppainen, director of research and innovation at Konecranes.

An inexpensive alternative to laser

The tracking of moving objects has been carried out for years. Traditionally, this has been done with laser cameras, but their costs can run up to tens of thousands of euros, support systems included, and the results they produce are still not always reliable. The visual SLAM technique is much less expensive than lasers, as the cameras used are based on a simpler technology. However, the data produced by the cameras contains gaps which must be covered with the help of algorithms.

“The hardest part is getting computer vision to recognise humans. This requires deep learning, a form of artificial intelligence,” Ruotsalainen says.

According to Ruotsalainen, developers of the visual SLAM technique in Finland and abroad have already for years been interested in deploying cheaper cameras. Even so, the method remains in need of much improvement.

“We have our work cut out for us in making the system reliable in challenging circumstances, such as industrial halls, where computer vision cannot necessarily identify any features to track. Reflections and fluctuations in brightness can also make the machine confused about its surroundings.”

The new study may in the future also benefit the development of self-driving cars, since the same questions related to positioning and observing the surroundings concern them as well.