GMDCSA-24: A dataset for human fall detection in videos

Area of Science:

• Gerontology
• Computer Science
• Biomedical Engineering

Background:

• The global elderly population is rapidly increasing, posing challenges for elder care and necessitating advanced fall detection solutions.
• Unintentional falls among older adults are a critical health concern, emphasizing the need for prompt detection and assistance.
• Remote monitoring systems are crucial for timely medical help following elder falls.

Purpose of the Study:

• To introduce the GMDCSA-24 dataset, designed to support the development of robust fall detection models for elders.
• To provide researchers with a versatile dataset for evaluating real-time fall detection systems in natural home settings.
• To facilitate research on the generalizability and robustness of fall detection algorithms against false positives from Activities of Daily Living (ADL).

Main Methods:

• The GMDCSA-24 dataset was created in three distinct natural home environments.
• Recordings included falls and various Activities of Daily Living (ADL) performed by four subjects under different lighting conditions (day and night).
• Low-resolution (0.92 Megapixel) webcam footage was utilized to ensure suitability for resource-constrained, real-time systems.

Main Results:

• The dataset comprises 81 fall and 79 ADL video clips.
• It captures diverse scenarios including varying clothing, lighting, and complex ADLs that might be misclassified as falls.
• The low-resolution videos are suitable for real-time processing without compression.

Conclusions:

• The GMDCSA-24 dataset is a valuable resource for advancing research in elder fall detection.
• It enables the development and testing of systems that are robust to real-world variations and complex daily activities.
• The dataset supports the creation of more reliable and efficient remote monitoring solutions for elder safety.