Poster
Feasibility on Detecting Door Slamming towards Monitoring Early Signs of Domestic Violence
IEEE/ACM IoTDI 2022.
Project Overview
Domestic violence remains a pervasive societal challenge, and early detection of warning signs can play a critical role in intervention and prevention. This project investigates the feasibility of using low-cost microcontrollers and TinyML to detect potential early warning signs of domestic violence and other anti-social behaviours within the home. The prototype focuses on classifying door closing events, specifically distinguishing between normal door closures and aggressive door slamming, using audio and accelerometer data captured by an Arduino Nano 33 BLE Sense. A TinyML model was trained and deployed directly on the microcontroller, enabling on-device inference without cloud connectivity. Under controlled laboratory conditions, the machine learning model achieved 88.89 percent accuracy in correctly classifying door slamming events, declining only marginally to 87.50 percent when assorted background noises were introduced to simulate realistic domestic environments.
These results show promising potential for supporting social workers and law enforcement agencies operating in social housing contexts, where continuous monitoring through conventional means is impractical. The approach prioritises privacy by performing all inference locally on the edge device, avoiding the transmission of raw audio data and ensuring that sensitive household sounds are never sent to external servers.
This work contributes to the emerging field of TinyML-based sensing for social good, exploring the boundaries of what resource-constrained embedded devices can achieve in sensitive monitoring applications. The project demonstrates that meaningful behavioural classification is possible on ultra-low-power hardware, opening pathways for discreet and privacy-preserving early warning systems in domestic settings.
Team
Osian Morgan
Hakan Kayan
Charith Perera