Human Daily Indoor Action (HDIA) Dataset: Privacy-Preserving Human Action Recognition Using Infrared Camera and Wearable Armband Sensors

Abstract

Human Activity Recognition (HAR) plays a vital role in applications such as healthcare, smart homes, and robotics, particularly in supporting the elderly. However, most existing HAR datasets focus on general human activities and are typically collected using RGB cameras in controlled environments with fixed angles—conditions that limit their real-world applicability. In this study, we introduce the Human Daily Indoor Actions (HDIA) dataset, specifically designed to capture natural indoor activities performed by elderly individuals. The dataset includes 48 daily actions, recorded using non-invasive infrared (IR) cameras and wearable armband sensors positioned at various angles to ensure diverse and realistic activity representation. The use of IR sensors enhances privacy, making the dataset ethically suitable for long-term monitoring. To demonstrate its utility, we implemented a fusion-based HAR model that integrates data from both IR and Inertial Measurement Unit (IMU) sensors. This model achieves strong activity recognition performance while minimizing the risk of identity exposure. By focusing on privacy-aware data collection and the daily routines of elderly individuals, the HDIA dataset offers a valuable resource for advancing real-world HAR research.