Real-world efficient fall detection: Balancing performance and complexity with FDGA workflow

• Published in: Computer vision and image understanding 2023-12, Vol.237, p.103832, Article 103832
• Main Authors: Zeng, Guotian, Zeng, Bi, Hu, Huiting
• Format: Article
• Language: English
• Subjects: Fall detection; Feature selection; Genetic Algorithm; Performance-speed trade-off; Skeletal information
• ISSN: 1077-3142; 1090-235X
• DOI: 10.1016/j.cviu.2023.103832

With the decrease in mobility, falls have become more prevalent and painful for both men and women. Most current approaches for fall detection in daily living employ either simple thresholds or complex networks. However, these methods suffer from an inability to consider the association among data or the disadvantage of excessive complexity in practical use. In this paper, we propose a compromise workflow called FDGA for fall detection that addresses these issues. Our approach first roughly determines the bounding box of the fallen target using the lightweight Fall-ROI extraction, and then further determines the precise skeletal position using a fine top-down skeletal estimator. Additionally, the genetic-algorithm-based classifier minimizes the effect of redundant features. Since no heavy network is required, the proposed workflow maintains an excellent performance-speed trade-off. We validate our approach using two fall detection datasets and a new large-scale fall detection dataset we created. The experimental results demonstrate the efficiency and effectiveness of our approach. [Display omitted] •Proposed fall detection workflow FDGA improves generalization in real-world scenarios.•GA-based feature selection optimizes skeleton-based attributes.•Built FDD15k dataset to compensate for limited fixed-angle datasets.•Systematic experimental make the proposed workflow more interpretable.