Skeleton-Based Posture Recognition for Home Care From Virtual Unmanned Aerial Vehicle

This article presents a novel approach for real-time posture recognition in monitoring scenarios, utilising a virtual camera simulated on a UAV within virtual environments. Leveraging the MediaPipe Pose library, key points of the body skeleton are extracted, focusing on a subset of 8 key points for...

Descripción completa

Detalles Bibliográficos
Autores: Bustamante, Andrés, Belmonte Moreno, Lidia María, Pereira, António, Morales Herrera, Rafael, Fernández Caballero, Antonio
Tipo de recurso: artículo
Fecha de publicación:2025
País:España
Institución:Consejo Superior de Investigaciones Científicas (CSIC)
Repositorio:RUIdeRA. Repositorio Institucional de la UCLM
OAI Identifier:oai:ruidera.uclm.es:10578/47916
Acceso en línea:https://onlinelibrary.wiley.com/doi/10.1111/exsy.70108
https://hdl.handle.net/10578/47916
Access Level:acceso abierto
Palabra clave:Heuristics
Home care
Posture recognition
Skeleton
Unmanned aerial vehicle
Virtual reality
Descripción
Sumario:This article presents a novel approach for real-time posture recognition in monitoring scenarios, utilising a virtual camera simulated on a UAV within virtual environments. Leveraging the MediaPipe Pose library, key points of the body skeleton are extracted, focusing on a subset of 8 key points for computational efficiency. Through the integration of heuristic algorithms based on physical proportions of the human body, the proposed methodology provides accurate estimations of three distinct postures: lying, standing, and sitting. This heuristic-based approach offers a computationally efficient alternative to traditional machine learning and deep learning methods, ensuring real-time performance and scalability. The efficiency of the framework is demonstrated through experiments that show its potential applications in various fields, including healthcare, virtual reality, and human-computer interaction. This approach achieved an average precision of 98.08% for virtual images. Success rates were 100%, 95.8%, and 98.9% for standing, sitting, and lying postures, respectively. Furthermore, the original classification model, which was tuned for virtual images, was tested on real images without any alteration to the parameter values. Its good performance demonstrates its potential for generalisation and application in diverse environments. Overall, this work contributes to the advancement of posture recognition technology, offering a versatile and accessible solution for posture analysis in dynamic monitoring environments.