Human activity recognition (HAR) is gaining importance in the development of smart personal protective equipment (Smart PPE). Yet, most solutions rely on body-worn devices or invasive helmet modifications, limiting adoption in industrial settings. We propose a noninvasive smart helmet with embedded motion and environmental sensors, fully integrated into a certification-compliant shell. To explore its applicability in realistic industrial conditions, we focused on constrained acquisition scenarios—characterized by low sampling rates, limited storage, and processing capacity. A multimodal dataset was collected using embedded triaxial accelerometers and barometric pressure sensors during operational tasks. Four classification pipelines were evaluated: random forest (RF), extreme gradient boosting (XGBoost), a wavelet-based RF (using horizontal concatenation of continuous wavelet transform, HC-CWT), and a deep learning (DL) pipeline combining a convolutional autoencoder with long short term memory (ConvAE-LSTM). Among these, XGBoost achieved the highest accuracy (up to 87.4%) and the most favorable balance between latency and computational efficiency, particularly with 8–15-s windows. This best-performing model was then deployed on the resource-constrained helmet system, confirming its suitability for on-board HAR in scalable occupational safety applications.
Design and Performance Evaluation of a Smart Helmet System for Human Activity Recognition in Worker Safety Applications
Schena E.;Setola R.;Massaroni C.
2026-01-01
Abstract
Human activity recognition (HAR) is gaining importance in the development of smart personal protective equipment (Smart PPE). Yet, most solutions rely on body-worn devices or invasive helmet modifications, limiting adoption in industrial settings. We propose a noninvasive smart helmet with embedded motion and environmental sensors, fully integrated into a certification-compliant shell. To explore its applicability in realistic industrial conditions, we focused on constrained acquisition scenarios—characterized by low sampling rates, limited storage, and processing capacity. A multimodal dataset was collected using embedded triaxial accelerometers and barometric pressure sensors during operational tasks. Four classification pipelines were evaluated: random forest (RF), extreme gradient boosting (XGBoost), a wavelet-based RF (using horizontal concatenation of continuous wavelet transform, HC-CWT), and a deep learning (DL) pipeline combining a convolutional autoencoder with long short term memory (ConvAE-LSTM). Among these, XGBoost achieved the highest accuracy (up to 87.4%) and the most favorable balance between latency and computational efficiency, particularly with 8–15-s windows. This best-performing model was then deployed on the resource-constrained helmet system, confirming its suitability for on-board HAR in scalable occupational safety applications.I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
