A Balanced big dataset for Sensor-Based Fall Detection: Enhancing Model Accuracy and Robustness

Abstract

Falls represent a critical challenge in healthcare, particularly for the elderly and those with limited mobility. They can cause severe injuries or deaths if not detected and addressed. Existing falldetection systems often struggle with lacking large, diverse, and balanced datasets; this limitation hinders the development of accurate and generalizable machine-learning (ML) solutions. This paper introduces a complete big dataset designed explicitly for video- and sensor-based fall detection, featuring 8,953 recorded activities, including 2,791 falls and 6,162 activities of daily living (ADL), collected from 29 diverse subjects. The dataset encompasses various fall scenarios—left, right, front, back, and complex cases such as attempting to sit on a chair or falling from elevated positions—along with ADL tasks such as walking, running, standing up from the ground, and driving. Each activity is recorded for 8 s at 100 Hz, yielding 800 data points per file. Including barometer-derived altitude-delta data significantly improves the performance of transformer-based models, raising accuracy from 97–98 % to more than 99.5 %. All 3,000 fall recordings were individually processed and non-matching patterns removed to confirm data quality, producing a clean and consistent corpus. Comparative experiments with existing datasets demonstrate superior detection accuracy and reduced false-positive rates, underscoring the robustness and reliability of our contribution. Overall, the proposed dataset provides the research community with a vital resource for advancing fall-detection systems and promotes the development of robust, deployable ML solutions for real-world healthcare applications.