Large language model-driven framework for inertial measurement unit-based worker activity recognition

Abstract

Inertial Measurement Units (IMUs) are widely used in wearable devices to detect worker activities, but current solutions often require multiple sensors and extensive labeled training data, limiting their practicality and applicability across diverse scenarios. This paper proposes a Large Language Model (LLM)-driven framework that recognizes worker activities from a single head-mounted IMU via unsupervised reasoning. Three meta-event principles are formulated and a video-IMU joint labeling tool is developed to extract meta-event features. An Activity Feature Recognizer is developed to identify motion characteristics, while K-Medoids clustering and autocorrelation functions are employed to quantify activity intensity and periodicity. Building upon these, an LLM-driven Agent network comprising a Supervisor, an Observer, and a Summarizer, is proposed to perform reasoning and activity understanding. Experiments achieved a Hamming distance of 2.62 for meta-event activity feature recognition and a 0.931 acceptance rate for Agent-inferred activity descriptions.