Fall Detection by Deep Learning-Based Bimodal Movement and Pose Sensing with Late Fusion

Area of Science:

• Gerontology
• Computer Science
• Biomedical Engineering

Background:

• Elderly fall detection is crucial but challenging.
• Single-sensor systems (IMUs, vision) often have high false positives and low accuracy in poor conditions.
• Existing methods struggle with real-time monitoring under varying light, including nighttime.

Purpose of the Study:

• To develop and evaluate a novel bimodal deep learning sensing framework for robust elderly fall detection.
• To improve accuracy and reduce false positives compared to single-modality systems.
• To ensure reliable real-time fall detection across diverse lighting conditions.

Main Methods:

• A bimodal deep learning framework integrating a memory-based autoencoder for inertial data and an attention-based network for visual data.
• Late fusion strategy at the decision level to combine sensor information.
• Experimental evaluation using a custom dataset of simulated falls and daily activities captured with waist-mounted IMUs and RGB cameras under dim lighting.

Main Results:

• The bimodal late-fusion system achieved a high F1-score of 97.3%.
• Significantly reduced false-positive rate (3.6%) compared to IMU-only (11.3%) and vision-only (8.9%) systems.
• Demonstrated robust performance in simulated fall detection under dim lighting conditions.

Conclusions:

• The proposed bimodal sensing framework offers a robust and accurate solution for elderly fall detection.
• The system's effectiveness is validated for real-time applications, even in low-light and nighttime scenarios.
• Combining inertial and visual data through deep learning significantly outperforms single-modality approaches.