Senso24
Simulating Safety: Using Synthetic Data to Train AI for Fall Detection
Skystrm Ltd, led by care industry veteran Justus Vermark, is transforming the way we respond to falls in elderly care environments. With an acute understanding of staffing limitations in the sector, Justus had already developed a promising AI capable of predicting fall events. But there was a barrier to progress: the lack of real-world data.
Acquiring large-scale, high-quality fall datasets—especially from vulnerable individuals—is practically and ethically impossible. Yet this data is vital to training AI systems that can make confident decisions in critical moments. To overcome this obstacle, Skystrm turned to CEMET with a question: could synthetic fall data, generated virtually, help train and improve the accuracy of their detection system?
CEMET’s Experimental Approach
Recognising the ambition and potential impact of the project, CEMET engaged in a focused, eight-week research sprint to test the viability of generating virtual fall data using the Unity game engine.
Over four structured sprints, the team designed a proof-of-concept system that could simulate human falls through a combination of animation blending and ragdoll physics. They captured detailed joint and movement data, exporting keypoint coordinates into structured CSV files for use in machine learning.
Each phase of development—ranging from rigging realistic animations, to automating data generation and labelling, to training and testing classifiers in JupyterLab—was carefully documented to ensure repeatability and a clear path for future scaling.
Machine Learning with Virtual Motion
The synthetic datasets were used to train a binary classifier and an anomaly detection model using TensorFlow. The results demonstrated that AI can, in fact, begin to learn the difference between “normal” behaviour and falls using entirely virtual data. This not only validated the conceptual model behind Senso24’s technology but also provided a testable framework for future real-world validation.
Further analysis explored how synthetic and real-world data (such as that captured by MediaPipe) might be bridged through transfer learning and normalisation techniques—laying out a roadmap for future development and investment.
A Clear Path Forward
While still in an early conceptual phase, this project has made it clear that synthetic fall data can provide an ethical, rapid, and scalable way of training fall detection models. It opens new opportunities for real-time care monitoring and could significantly improve safety in residential care settings.
“Working with CEMET has been an incredible experience for Skystrm Limited, particularly on our Senso24 project... Their support significantly enhanced our ability to attract further investment and positioned us to expand into new markets.”
Impact and Next Steps
The collaboration between CEMET and Skystrm Ltd has laid the groundwork for a powerful new approach to fall detection in care settings. By demonstrating that synthetic data generated within Unity can be used to train reliable AI classifiers, the project validated the core concept behind Senso24’s innovation. The entire development process was carefully documented, ensuring that future iterations can build on this foundation with clarity and consistency.
Crucially, the project has not only delivered a working proof of concept but also clarified a roadmap for integrating real-world data through techniques like transfer learning and data normalisation. With this groundwork in place, Senso24 is now in a strong position to seek further investment, scale development, and begin real-world testing. The results also highlight a compelling path forward for using ethical, virtual simulations to support vulnerable individuals—without compromising on safety, privacy, or dignity.
CEMET is proud to continue to support Skystrm through the development or this new service. To find out more about their latest work visit their website
This project is jointly funded by the UK Government’s Shared Prosperity Fund and Cardiff Capital Region (CCR) the Academic-Industry Partnerships programme, part of the Cluster Development and Growth Programme.
