Fall Detection Based on Computer Vision: A Systematic Mapping Study.

Area of Science:

• Computer Vision
• Gerontology
• Biomedical Engineering

Background:

• Falls are a major cause of injury and death in older adults.
• Video-based computer vision (CV) systems offer a promising solution for fall detection.
• Existing research lacks a consolidated overview of current CV fall detection methodologies.

Purpose of the Study:

• To systematically map the field of vision-based fall detection in video.
• To analyze trends in algorithmic approaches, performance, and efficiency reporting.
• To identify research gaps hindering clinical translation and reproducibility.

Main Methods:

• Systematic mapping of 433 primary studies published through 2025.
• Utilized five databases with explicit eligibility criteria and independent dual screening.
• Employed a PRISMA-inspired workflow for evidence synthesis.
• Characterized approaches using a three-level taxonomy: Feature Engineering, Deep Learning, and hybrid models.
• Quantitatively analyzed algorithmic components, performance metrics, and efficiency data (FPS, latency).

Main Results:

• Deep Learning pipelines, especially Convolutional Neural Network (CNN)-based backbones, are predominant.
• Hybrid models remain prevalent, with Transformers/Attention architectures showing rapid recent adoption.
• Despite claims of real-time performance, efficiency metrics and hardware specifications are inconsistently reported.
• This inconsistency limits the reproducibility and clinical translatability of CV fall detection systems.

Conclusions:

• The field is dominated by Deep Learning and hybrid approaches, with emerging Transformer/Attention models.
• Inconsistent reporting of efficiency and hardware details is a significant barrier to practical deployment.
• Further standardization in reporting is crucial for advancing reproducible and clinically translatable fall detection technologies.