Multi-scale motion perception fall detection algorithm based on video swin transformer

Fall detection is a prominent subject in healthcare. Advancements in modern monitoring and deep learning have sparked significant social interest in visual fall detection. Despite the success of various deep learning methods in video fall detection owing to their superior feature extraction capabili...

Ausführliche Beschreibung

Gespeichert in:
Bibliographische Detailangaben
Veröffentlicht in:Signal, image and video processing Jg. 19; H. 10; S. 800
Hauptverfasser: Cai, Xi, Wang, Xiangcheng, Jiao, Yin, Chen, Yinuo, Han, Guang
Format: Journal Article
Sprache:Englisch
Veröffentlicht: London Springer London 01.10.2025
Springer Nature B.V
Schlagworte:
Accuracy
Algorithms
Computer Imaging
Computer Science
Deep learning
Design
Efficiency
Fall detection
Falls
Feature extraction
Image Processing and Computer Vision
Machine learning
Modules
Motion perception
Multimedia Information Systems
Neural networks
Original Paper
Pattern Recognition and Graphics
Signal,Image and Speech Processing
Surveillance
Vision
Deep learning
Computer vision
Video swin transformer
Fall detection
ISSN:1863-1703, 1863-1711
Online-Zugang:Volltext
Tags: Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
Beschreibung
Zusammenfassung:Fall detection is a prominent subject in healthcare. Advancements in modern monitoring and deep learning have sparked significant social interest in visual fall detection. Despite the success of various deep learning methods in video fall detection owing to their superior feature extraction capabilities, they still encounter challenges in analyzing long-range or short-range spatiotemporal correlations. Taking this into account, a multi-scale motion perception fall detection algorithm based on video swin transformer is proposed in this study. Our proposed method employs video swin transformer as the backbone to fully model the global and local spatiotemporal information from videos and optimizes the backbone with two integrated modules. On one hand, we design a multi-scale motion information aggregation module to overcome the difficulty of the model in focusing on key multi-scale motion features. On the other hand, we propose a token pruning module to reduce the computational cost by pruning redundant temporal tokens. Experimental results demonstrate that the proposed algorithm exhibits promising outcomes, with an accuracy of 96.11% and 97.05% on the Le2i and UR fall detection datasets, respectively, thus outperforming some existing advanced algorithms.
Bibliographie:ObjectType-Article-1
SourceType-Scholarly Journals-1
ObjectType-Feature-2
content type line 14
ISSN:1863-1703
1863-1711
DOI:10.1007/s11760-025-04358-3