Community Spotlight: Michael Dubbeldam

Building FlowCube: Inside vision AI for traffic management

Every month, we spotlight a member of our community. At KickstartAI, we believe that meaningful progress starts with individuals — engineers, researchers, founders, and practitioners — who are applying AI in bold, practical ways. Through our Community Spotlight series, we celebrate the people shaping the future of AI in the Netherlands: their work, their challenges, and the lessons they’re learning along the way.

This month, we’re featuring Michael Dubbeldam, Domain Architect at Technolution — the company behind FlowCube, a camera-based traffic sensor with embedded deep learning models. From solving real-world mobility challenges in Copenhagen to pioneering vision AI, this is his story.

Michael Dubbeldam started out as a software engineer and, later on, moved into software architecture. He joined Technolution in 2005, where he’s now leading the AI competence group. His interest in AI was first sparked by a traffic management project for the city of Copenhagen in 2017. 

“The city wanted a system that could perform as well for cyclists as it did for cars,” Michael recalls. “They needed to evaluate how well the streets were serving cyclists and collect information on how much time cyclists need to go from one side of the city to the other via a particular route.” 

At that time, the only way to gather that data was by physically recording videos and then manually analyzing the footage to track individual cyclists. 

Naturally, that process took an enormous amount of time. “At that point, we thought: if only there was a system that could do something like that. If only we could use a camera to see what we see.” 

Michael explains the rationale behind that idea. 

“Traditional traffic management systems relying on traffic light input provide limited information,” he says. “But as humans, going into the field and seeing what’s happening, we can perceive so much more.” 

A camera could potentially replace human eyes, but there was a challenge: enabling a computer to provide and translate all the necessary information. 

Michael believed that vision AI could be the answer. 

“I thought that with AI, computers can see the same things or understand the same things as humans. So you don’t need to adjust the world to computers; instead, computers can adjust to the human world.” 

“That’s why I really got into AI. I wanted to understand how it works, build my own models, and push the boundaries of what’s possible.”

The Copenhagen challenge marked the first vision AI project for Michael and the Technolution team. It also served as the inspiration for a new product journey. 

FlowCube: from concept to product 

Michael’s first foray into vision AI gave birth to the concept of  FlowCube —  a camera-based traffic sensor with embedded deep learning models that measure traffic flows of vehicles, bicycles, and pedestrians.

“When we first started, it wasn’t a cube,” he says. Initially, the team created a device with a camera that processed images locally, converting them into data. The system would send information in the form of “I saw a cyclist with a blue coat at this location at this moment,” Michael explains. That way, it could analyze traffic without storing or transmitting any raw images. 

“Over time, we designed a nice cube and kept making the system smarter and more efficient,” he says. 

It took about two years to reach the minimum viable product stage. “Once you’ve had that, you think that you’re done and can take on the world,” Michael adds. “But as you build, you learn more about what you need to do. We’ve kept refining and improving the system ever since, and I expect it will be even better five years from now.” 

Today, FlowCube comes as a box measuring 13x13x13cm. Inside, there’s an inexpensive commodity camera, a computer, an Nvidia processor, and a 4G mobile connection. 

Building with Siamese models 

FlowCube captures images and is able to detect objects and people and map their movement in real time. 

“We get information like: there’s a cyclist with a blue coat and a bag on the bike at this location and time,” Michael says. The device then converts this type of data into what the team calls a “match code” — a statistical representation of an object’s appearance. 

These match codes are sent to a cloud server, where data from multiple FlowCube devices is combined to reconstruct city-wide mobility patterns. This way, the system provides a complete picture of traffic across all road users, while preserving the privacy of individuals. 

“We don’t track people,” Michael emphasizes. “It’s about statistical patterns and comparing likelihoods.” For example, that a cyclist at point AI is statistically similar to another one at point B. 

To achieve this, the team uses Siamese models — trained not to classify objects, but to assess how alike two inputs are. In this case, not whether two people are both cyclists, but whether they might be the same individual seen in different locations. 

That’s useful for what’s called “person re-identification,” where the goal is to determine how much two people resemble each other. 

The system is trained using labeled images, so it learns what features actually help distinguish individuals. “For example, it might learn that background or the fact that someone is on two wheels isn’t helpful, but clothing color or a backpack might be.”

Michael’s team is also working to make the system’s decisions more interpretable. “We’re now exploring how we can name those distinguishing features explicitly — like ‘white sneakers’ or ‘blue coat’ — so the reasoning is understandable not just to machines, but to people too.”

For model training, Technolution creates its own datasets. “For the re-identification models, for example, we have a dataset of approximately 250,000 images and 15,000 identities,” Michael says. “We’ve also needed to adapt our training data to the local context.” For instance, vehicles in the US, include pick-up trucks and yellow school buses — which are uncommon in Europe. In Nordic countries, cyclists typically wear helmets unlike in the Netherlands.

Challenges 

Michael identifies three main challenges in building and developing FlowCube. The first one is hardware: building a device that’s robust, scalable, easy to manufacture, and consistently reliable in the field. 

Another difficulty is creating models that generalize well across different environments. Something that works in one place may not work elsewhere without retraining.

Finally, the third major challenge lies in the development of the Siamese models. “Most research on the technology of person or object re-identification works with closed-world benchmarks, where a match is guaranteed. You simply have to find the best one.” 

But in real open-world environments, this isn’t the case. “You don’t know if a match exists,” Michael explains. 

This gap between closed- versus open-world applications meant that the team couldn’t rely on existing models or benchmarks. “We had to define our models, layers, cost functions, and training to get them to do what we needed them to do.” 

AI’s role in multimodal traffic management 

For Michael, the biggest benefit AI brings to traffic management is access to human-level information. Traditional sensors can’t distinguish a bus from a car based on visual cues. “Vision AI can,” he emphasizes. 

AI-based sensors can make smarter decisions and optimize the overall traffic management system — for example, by adjusting traffic lights to prioritize buses. This optimization is particularly important for multimodal countries, with a large number of cyclists and pedestrians, like the Netherlands. 

Equally important is to comply with regulations like the GDPR in order to increase both customer and public trust. 

“Images never leave the device — they are processed on the edge and then discarded. We don’t store identifiable information,” Michael says. 

“But we also need to involve the cities or municipalities we’re working with in our processes, so they understand our privacy safeguards.” 

Michael joined the KickstartAI community only recently, but already sees value. 

“There aren’t many people who are deeply in this field and who are bringing lab work to reality. So it’s good to see how others manage the same things we struggle with. To me, that’s really educational. And I’m reading a lot that I find very interesting.” 

💡 Want to learn more about the KickstartAI community and our mission to accelerate AI adoption in the Netherlands? Check out our website, listen to our podcast, or follow us on LinkedIn!