Neuromorphic Vision-Based Fall Localization in Event Streams With Temporal-Spatial Attention Weighted Network.

Area of Science:

• Computer Vision
• Biomedical Engineering
• Gerontology

Background:

• Falls are a major cause of accidental death in the elderly, necessitating advanced detection methods.
• Existing fall recognition systems using wearable or standard vision sensors pose privacy risks and struggle with localizing falls in lengthy recordings.
• Bioinspired vision sensors offer a privacy-preserving alternative with high dynamic range, responding to brightness changes asynchronously.

Purpose of the Study:

• To propose a novel framework for precise temporal localization of falls using bioinspired vision sensors.
• To address privacy concerns and limitations of existing fall detection methods.
• To develop a system capable of accurately identifying fall events within long video streams.

Main Methods:

• Utilized a dynamic and active-pixel vision sensor (DAVIS) camera for its privacy-preserving and high dynamic range capabilities.
• Developed an adaptive temporal window conversion mechanism for improved event data representation.
• Implemented an event temporal density score and the 1D-watershed algorithm for efficient proposal generation, combined with temporal and spatial attention mechanisms for fall modeling.

Main Results:

• The proposed framework demonstrated precise temporal localization of simulated falls.
• Achieved state-of-the-art performance in fall detection accuracy and localization.
• Validated through experiments with 30 volunteers performing simulated falls.

Conclusions:

• The bioinspired vision sensor-based framework offers a privacy-preserving and accurate solution for fall temporal localization.
• This technology has the potential to significantly enhance safety monitoring for elderly individuals living alone.
• The adaptive temporal window and event-driven proposal generation are key innovations for effective fall detection.