Validation of Virtual Human Motion Sensors in a Digital Twin Environment Using LLM-based Activity Generation

Abstract

With the growing use of IoT devices in smart homes, accurate human activity recognition (HAR) is crucial to enable context-aware services. While cameras and microphones offer rich contextual data, they raise privacy concerns in residential environments. Passive infrared (PIR) motion sensors offer a privacy-preserving alternative, but their detection performance heavily depends on sensor placement, which is typically determined empirically. This paper presents a digital-twin-based simulation framework for evaluating PIR sensor placement in indoor environments and supporting data-driven optimization in future work. Our system reconstructs a 3D virtual home with virtual PIR sensors and LLM-based agents. By comparing virtual and real sensor logs, we confirmed that the virtual sensors can approximate real sensors' detection responses with moderate yet meaningful agreement (mean precision = 0.62, mean recall = 0.72, accuracy = 0.87).