A Lightweight Skeleton-Based Fall Detection Framework Using Multi-Dimensional Attention Mechanisms

Abstract

Fall detection is a critical task in intelligent healthcare and smart monitoring systems, particularly for elderly care, where timely and reliable detection can significantly reduce the risk of severe injuries. In this paper, we propose a lightweight fall detection framework based on human skeleton sequences and multi-dimensional attention mechanisms. Instead of relying on raw RGB information, the proposed approach leverages skeleton-based representations to reduce sensitivity to background clutter, illumination variation, and appearance differences.

The framework consists of four main stages: skeleton keypoint extraction and preprocessing, spatial feature encoding using a Coordinate Attention enhanced Transformer, temporal motion modeling via a Temporal Attention mechanism, and final classification. The Coordinate Attention module enables direction-aware spatial modeling of human joints, while the Temporal Attention mechanism adaptively emphasizes critical motion phases associated with falls, such as sudden posture collapse and rapid descent.

Extensive experiments are conducted on two public benchmark datasets, namely the University of Rzeszow Fall Detection Dataset (URFD) and the Multiple Cameras Fall Dataset (MCFD). The experimental results demonstrate that the proposed method consistently outperforms several representative state-of-the-art fall detection approaches, achieving higher precision, recall, and F$_1$-score. In particular, the significant improvement in recall highlights the effectiveness of the proposed model in reducing missed fall detections, which is crucial for safety-critical applications.

Furthermore, the proposed framework is designed with computational efficiency in mind. By employing a compact Transformer architecture and lightweight attention modules, the system achieves real-time inference on CPU-based platforms, making it suitable for deployment in resource-constrained environments such as smart homes and elderly care facilities. Overall, this work demonstrates that combining skeleton-based representations with efficient spatial–temporal attention mechanisms provides a practical and reliable solution for real-world fall detection.