Fake data for real diagnosis
“Our scientific proposal is to add neural networks to EPSI’s rule-based algorithms,” says Frédérick Benaben. The advantage of these networks is that the more data you provide, the more precise and efficient they become. “The issue is around feeding them,” says the researcher. “Because EPSI works on sensitive sites, where it is complicated to share data.” IMT Mines Albi’s first contribution will therefore be to generate synthesized data to train the neural networks: first on “undeformable” objects (drones, vehicles, etc.), then on “deformable” objects (humans, animals, etc.) that are more complex to interpret.
One way of “manufacturing” data for an undeformable object is to record the object’s radar signature from several angles, like a photograph, generate a fictitious trajectory and associate signatures corresponding to the radar’s points of view of this trajectory. A large amount of fake data can then be obtained by digitally simulating combinations of possible trajectories and radar signatures. Another way is to perform “augmentation”, i.e. to generate data using existing data, “like a photo manipulated from several angles, zoomed in or shifted to one side, producing other photos”.
When it comes to representing deformable objects, scientists are exploring the idea of adding up diagnosed movements: does the signature of a person walking correspond to combining the signatures of a moving arm, leg, two legs, etc.? “In principle, there is little variability between individuals, so we hope that for two people doing the same thing, the signature will be almost identical.” EPSI’s expertise will be particularly valuable at this stage.
Specialized neural networks
Rather than creating a single general-purpose detection system, which would be complex to train, the RadaR-IO research teams plan to develop several types of neural networks, with different and complementary skills. Using multiple neural networks of varying levels of specialization in parallel would increase diagnostic accuracy.
There is just one snag: “we could imagine developing a neural network specializing in small animals or vehicles, for example. The problem is, we don’t know how they would work or what their classification rules would be!” says Frédérick Benaben. “So it’s hard to say whether size or the way something moves would be a criterion for classification.” This training stage also raises the question of the level of precision required: is it important for the system to be able to differentiate between a rabbit and a cat, for example?
The synthesized data will then be fed into the various neural networks developed, and the challenge for the scientists will be to understand or interpret what comes out of it, under the guidance of Aurélie Montarnal, also a researcher at IMT Mines Albi. The effectiveness of the systems developed will have to be compared with the diagnostic tools currently available to EPSI.
Time for action!
Once the radar data has been retrieved and interpreted, RadaR-IO’s aim is to offer a complementary range of recommendations, or even measures triggered immediately, depending on what is detected. “What interests us, beyond simple diagnosis, is automation, semi-automation and proposing measures to be taken. All this is also in our field of expertise,” says Frédérick Benaben. Measures to assist with crisis management must logically take into account EPSI’s market and its customers’ needs.
An alert on a site under surveillance may involve a simple, straightforward procedure: launching ultrasound repellers if birds are detected at an airport, or establishing a security perimeter around a work of art, for example. “On the other hand, in the case of a penitentiary, the range of options is much more elaborate. Just because the system identifies a person crawling doesn’t mean they should automatically be shot at,” says the researcher. “These buildings are subject to a set of rules to which our tools must also conform.”
These rules are generally defined in dense, complex documentation, but AI tools can extract data from them that can be used as instructions. Said instructions would then be fed into radar diagnostic operation and response mechanisms. “We’re not there yet, but we’re exploring these areas as we think they would be really interesting to integrate in this context,” says Frédérick Benaben.