TExNet: long short-term fast fall detection based on attention enhancement and self-adaption

Area of Science:

• Computer Science
• Artificial Intelligence
• Biomedical Engineering

Background:

• Falls represent a critical safety concern, necessitating swift and precise detection systems.
• Existing fall detection methods often struggle with the discrepancy between simulated and real-world data.

Purpose of the Study:

• To propose the Time Exploration Network (TExNet), an attention-enhanced model for accurate and rapid fall detection.
• To address the simulation-to-real gap in fall detection by integrating multi-branch timing and classification.
• To enhance model robustness against data distribution variations and improve action understanding.

Main Methods:

• Developed a two-branch adaptive fusion framework combining Convolutional Neural Network (CNN) and Transformer architectures.
• Incorporated dilated convolution and time series positional decomposition for improved temporal correlation.
• Utilized an Invariant Risk Minimization (IRM) inspired loss function and a data self-conditioning module.
• Employed a fine-tuning strategy with a pre-trained framework and few-shot learning for downstream tasks.

Main Results:

• The fine-tuned TExNet model achieved a recall of 92.16%, demonstrating strong adaptability to new data distributions.
• Experimental results confirmed the superiority of TExNet over existing fall detection approaches.
• The model effectively captures both local and global dependencies and enhances temporal understanding.

Conclusions:

• TExNet offers a significant advancement in fall detection technology by effectively handling real-world complexities.
• The proposed methods improve model generalization and robustness, crucial for safety-critical applications.
• TExNet demonstrates high performance and adaptability, paving the way for more reliable fall monitoring systems.