Wide range of video applications
Companies are using video more and more often, in a wide range of applications: from surveillance and other situations in which you want to monitor any deviation from a normal situation to inspection and investigation of objects. In most companies, viewing and inspecting these images is still a matter of manual work. Frank is training machine learning algorithms to recognize video images. The purpose: fully automatic detection of occurrences of a particular situation. This will allow people to focus on that part of their work that requires their knowhow. Frank: “The time savings are huge, and what’s more, we’re making the work of the staff who need to analyze those images a lot more pleasant.
Fully automatic damage detection
Frank was involved in a project for BAM, during which he trained a self-learning model to recognize various types of asphalt damage. He describes: “BAM is using scan-vehicles to take photos of the road surface. In the past, those photos were analyzed by inspectors. However, most photos don’t show any damage, and it becomes ultra-boring work, but at the same time, you have to remain very focused. We trained an AI model to recognize eight types of damage fully automatically. The first algorithm could already indicate, with certainty, that 80 percent of the images did not show any damage. For the remaining 20% the algorithm proposed further analysis. Those images were then analyzed by inspectors and we used their feedback for further training of the model. Our goal for the future is that we will be able to detect damage in 99% of the cases fully automatically.”
“We’re training AI models so that they can automatically classify video images.”
Outgrown its infancy years
Frank gets inspired when he is showing the business case to customers. “A lot of companies believe that this technology is still in its infancy stage, but nothing is further from the truth. The police have been using AI for years in investigations, for instance for automatic recognition of license plates. And hospitals use it for analyzing radiology images. A welltrained algorithm increases the quality of the analysis and reduces the time required for analyzing the images. Which means that better decisions can be made in less time. The human capacity is no longer the limiting factor. That higher speed can save human lives in diagnostics or investigations. Moreover, application of AI can lead to new revenue models in the inspection and investigation of objects.”
Solving puzzles with data sets
As a data scientist, it’s Frank’s work to create transparency in large data sets which will subsequently lead to more insight. “I’m always solving puzzles: if I add another source to my data set, will that result in new insights? Or can I use an existing source for other applications? For example, can we use the photos that BAM is taking of the roads also for assessing whether traffic situations are clear and/or whether traffic signs are placed in clearly visible and logical locations? And are there any opportunities for other revenue models due to images being analyzed much faster? This is how I often present ideas for new applications to customers. That’s great, because despite the fact that the core of my work is very technical, I’m constantly acting on the common ground that is shared between business and IT.”
