Sensing with signals: Making environments smarter using radio waves and machine learning
PhD researcher Bram van Berlo used radio signals and AI to sense human activity without cameras or wearables, enabling more accessible and unobtrusive technology.
From health monitoring to smart living spaces, there is a growing interest in systems that understand human presence and behavior without requiring people to carry or wear sensors. PhD researcher Bram van Berlo explored how radio waves, which are already used for wireless communication, can also be used to sense what is happening in a room. He defended his thesis on Monday, July 7.
Rethinking how we sense
Traditional sensing devices such as smartphones, watches, and fitness trackers require user participation and may not be suitable for all individuals, especially those with certain medical or psychological conditions.
Radio waves offer a promising alternative in sensors. They reflect and scatter when they encounter people or objects, and these changes can be measured.
However, interpreting those measurements is complex. Traditional models rely on human expertise to extract meaningful features from signal data. Van Berlo used machine learning to discover those patterns automatically, which reduces the need for expert input and improves scalability.
A major challenge remains though. Machine learning models often perform well in the environment they are trained in, but their accuracy can decrease when used in a new space or with new people. This is due to differences in signal behaviour, known as distribution shifts.
New techniques to improve reliability
To address this, developed a method called mini-batch alignment. This training approach helps the model focus on signal features that remain stable across different environments. It improves generalization without increasing training time or requiring structural changes to the model.
He also experimented with using environmental labels during pre-training to make models more adaptable to new settings. By organizing data so that training batches represent a variety of conditions, he reduced conflicts during learning and improved performance.
Real-world applications
Van Berlo鈥檚 research focused on two main applications. First, he looked at activity recognition, where the system detects actions such as walking or standing based on how centimeter-wave signals interact with the environment.
Second, he explored path blockage prediction, where the system anticipates when a person or object will interfere with a signal path. This is relevant for smart buildings, wireless networks, and assistive technologies.
He led a large data collection effort to understand how signal conditions affect model accuracy. He identified key factors for reliable sensing, including signal resolution, data balance, and the relationship between signal segments and the target activity. He also used feature attribution techniques to reveal the signal elements that contribute most to accurate predictions.
PhD researcher Bram van Berlo. Photo: Angeline Swinkels
Broader impact
The research conducted by Van Berlo contributes to the development of intelligent environments that are more inclusive, adaptive, and easy to use.
His techniques can help machine learning models perform reliably in new conditions without retraining; opening doors to smarter healthcare, communication, and safety systems.
-
Supervisors
Nirvana Meratnia, Tanir Ozcelebi
Written by
More on AI and Data Science
Latest news
