“People don’t know how big of a problem it really is”, says student Reinis. “It is terrible in post-conflict regions where these mines have been used in warfare. And in the end, most of the people who end up suffering are civilians: children playing, farmers working, citizens walking around in those areas. And just like that, they could lose a limb.”
Though technological solutions are advancing, to this day, demining is still a man’s job. “It’s still usually done walking around with a metal detector. And obviously, that’s very risky, and also extremely expensive and time-consuming. Every time it beeps, you have to assume it’s a landmine and be extremely cautious – only to realise that it’s just a beverage can.”
DroneTeam Twente
The third-year Mechanical Engineering students realised that there is a great demand for automated demining technologies. Being members of the DroneTeam Twente, they share a passionate interest in drone technologies – and how they can be used for humanitarian aid purposes. “The challenge that we’re participating in with the DroneTeam Twente is to design a medical aid delivery drone. You’re being rated on all kinds of aspects: design, flying technique, battery efficiency, how much cargo you can transport and even marketing and business proposition.”
Trying to find a subject for their thesis, they came in touch with Alberto Martinetti, associate professor in Humanitarian Engineering at the University of Twente – who is also involved in setting up a new Master’s programme in Humanitarian Engineering. “He told us about a UT-/VU-funded demining project he was involved in, and it seemed like the perfect topic for us”, says Benjamin. “And the best part was that we were able to work on our thesis while still being part of the DroneTeam, so we could perfectly integrate these two. This way, we didn’t suffer any study delay.”
Machine learning model
Sitting in the workspace of the DroneTeam, Reinis holds up a 3D-printed drone, with a thermal camera attached to it. “This is one of the drones that we developed ourselves.” “Reinis focused more on the development of the drones, whereas I concentrated more on the computer science side of it”, continues Benjamin. “Basically, we trained a machine learning model to recognise a PFM-1, a Russian landmine that is scattered out of helicopters. These kinds of mines are commonly used and remain on the surface. Ukraine is full of them, for example.”
This kind of research is not new, although Benjamin and Reinis had a clear objective with their thesis. “Mostly, research is about how to reach the highest accuracy in detecting mines. But we aimed to research: how low in accuracy can you go? How quickly can the drone fly before the image becomes too blurry? Because you want to be efficient, of course. So, our question was: how can we make it accessible, cheap and standardised? It would be great to have systems that are not unreasonably overpriced and can be widely implemented in all parts of the world.”
While working on their thesis project, there were quite some challenges to overcome. For example in building an efficient drone, says Reinis. “How do you maximise the range? Of course, you can add more batteries: but that will take more power because it becomes heavier. We’ve used specialised printing methods and carefully picked materials to make the drones as lightweight as possible.”
Using wax instead of TNT
To be able to test the machine learning model, the students needed to get creative as well. After all, you will not perform tests with an actual landmine. That’s why they built their own replica of a PFM-1 landmine with a 3D printer. “And instead of the explosive TNT, we used paraffin wax. This is the kind of wax that is also used for candles, and it has the same thermal properties as TNT”, says Benjamin. “And then we took a whole lot of thermal pictures of these landmines, either on the ground, in the grass, in the sand, some of them were even partially buried. In the end, we wanted to train the model to see it, even though it’s buried under the soil”, adds Reinis.
With over 800 images of their own-developed mine being analysed by the machine learning model, they started testing on various locations, by spreading 10 mines replicas across a field and also other (metal) objects that should not be recognised as a mine. “We had our fair share of crashes, but eventually it worked really well, actually! The drone was able to recognise most of the mine replicas while ignoring the other objects. By the end, we got about more than 90% accuracy.”
Now that the thesis is coming to an end, the students look back on a successful project. “We realise that it will never be 100% accurate, and there will still be the need for that person on the ground with a metal detector. But say, a whole countryside is potentially contaminated, then drone technologies can really narrow down the area and set priorities on where to start demining actions.”
