Visual object tracking is a field in which Professor Michael Felsberg and his co-workers have led scientific development since 2013: their prolific scientific publications have not only been awarded several prizes, but also been cited by many other researchers. The newly established company Singulareye is based on the research results.
Interest in visual object tracking has exploded in recent years as a consequence of increasing interest in autonomous vehicles, robotics, artificial intelligence, and machine learning. The technology has advanced rapidly and one reason for this is something as unusual as a competition, the Visual Object Tracking (VOT) Challenge, arranged each year since 2013.
“The competition gathers the most renowned researchers within the field at leading universities throughout the world”, says Michael Felsberg, professor in computer vision at Linköping University.
He has been a member of the competition organising committee for some years now. Michael Felsberg won the prestigious competition in 2014, together with his doctoral student at the time, Martin Danelljan, Fahad Khan, a co-supervisor and researcher in the division, and Gustav Häger, at the time an undergraduate, now doctoral student. As a continuation of his degree project, Martin Danelljan had developed a method in which the computer follows an object within a defined box, known as a “bounding box”, in the image.
The computer learns the appearance of the object in the box in different conditions, both its colour and shape, with the aid of features in the image – edges or colours. The calculations are then rapidly performed using carefully chosen algorithms. The group had in their work removed an important stumbling block within object tracking: the problem of scalability – a moving object changes its apparent size and shape as it recedes from the camera.
The group published the work as open-source code, which led to thousands of citations in a very short period. The research carried out in the group is partially financed by ELLIIT, and has subsequently been awarded several prizes.
When Martin Danelljan defended his doctoral thesis in 2018, his work had already received over 2000 citations, and today the number is over 9000. In the past two years, the research has taken a further step.
“The research is now, however, starting to increasingly use a technique known as ‘segmentation’ in which pixels in the image are classified as belonging to different objects. In classification the computer should be able to distinguish, for example, a sleeping dog from the sofa on which it is sleeping. To put it simply, segmentation means doing the same thing at pixel level”, says Michael Felsberg.
Development in the field has increasingly come to be dominated by Chinese researchers, and they won all prizes except one in the most recent edition of the VOT Challenge.
“The Chinese have large resources for research in the field, in particular for improving facial recognition. This is an accepted technology in China, but is ethically problematic in the western world,” says Michael Felsberg.
Other areas of application for visual object tracking are both broader and less controversial. Examples include autonomous vehicles that can discover and follow pedestrians and bicycles in traffic, flying vessels that search for people in need of help in disaster areas, and mobile phone cameras that can focus automatically on a face in the image (if any).
“In a few years we will probably have a function in all mobile phones where we can record video and switch backgrounds while recording”, says Michael Felsberg.
The objectives of the research, however, are higher than that.
“We construct our world-view primarily on what we see and experience, and robots that are to function together with people must have the same ability. It must be possible for AI systems to learn from data without without complete annotation, and based on only few examples (one-shot learning). We cannot expect that all the data processed by artificial intelligence will first be created by a human and then learnt by a system. We must be able to define a starting point and direction, and then leave the learning process to proceed unsupervised”, he says.
Martin Danelljan, who is today (2020) a postdoc at ETH in Zürich, founded the company Singulareye in 2018, together with two other doctoral students. In addition to world-leading knowledge in visual object tracking, the company offers consultancy services with expertise in deep learning and AI. Its largest customers are found in the automotive industry.