Publicación: Desarrollo de un modelo de reconocimiento de acciones en procesos industriales
Cargando...
Fecha
2023-09
Autores
Editor/a
Director/a
Tutor/a
Coordinador/a
Prologuista
Revisor/a
Ilustrador/a
Derechos de acceso
Atribución-NoComercial-SinDerivadas 4.0 Internacional
info:eu-repo/semantics/openAccess
info:eu-repo/semantics/openAccess
Título de la revista
ISSN de la revista
Título del volumen
Editor
Universidad Nacional de Educación a Distancia (España). Escuela Técnica Superior de Ingeniería Informática. Departamento de Inteligencia Artificial
Resumen
El reconocimiento de acciones humanas ha sido una problemática de gran interés en las últimas décadas debido a su gran cantidad de aplicaciones. Entre ellas, la aplicación a entornos industriales, que hasta ahora había sido más limitada debido a las dificultades de recolección de datos, brinda grandes oportunidades en cuanto a un aumento de la trazabilidad, control de seguridad, de protocolo y rendimiento del operario. El conjunto de datos OpenPack ofrece gran cantidad de datos para clasificar acciones en ambientes industriales recogidos a partir de múltiples sensores en un entorno controlado y propone un concurso para desarrollar un sistema capaz de reconocer la acción realizada por un operario. En este trabajo se ha pretendido utilizar una serie de modelos combinados multimodales que procesan los datos de sensores inerciales y cámaras visibles para clasificar cada acción realizada por el operario alcanzando resultados equiparables a los del top de competidores que han participado en el concurso propuesto por los creadores del conjunto de datos. Para ello se han aprovechado las características de las redes neuronales convolucionales y redes neuronales recurrentes para datos provenientes de sensores incerciales y redes neuronales basadas en grafos para la etapa visual a través de estimación de pose con keypoints dados.
Human action recognition has been a topic of great interest in recent decades due to its wide range of applications. Among them, its application in industrial environments, which had been more limited until now due to data collection difficulties, offers significant opportunities in terms of increased traceability, safety control, protocol adherence, and operator performance. The OpenPack dataset provides a wealth of data for classifying actions in industrial settings, collected from multiple sensors in a controlled environment, and proposes a competition to develop a system capable of recognizing the actions performed by an operator. This work aimed to employ a series of multi-modal combined models that process data from inertial sensors and visible cameras to classify each action performed by the operator, achieving results comparable to those of the top competitors who participated in the competition proposed by the dataset creators. To achieve this, we leveraged the features of convolutional neural networks and recurrent neural networks for data from inertial sensors, and graph-based neural networks for the visual stage through keypoint-based pose estimation.
Human action recognition has been a topic of great interest in recent decades due to its wide range of applications. Among them, its application in industrial environments, which had been more limited until now due to data collection difficulties, offers significant opportunities in terms of increased traceability, safety control, protocol adherence, and operator performance. The OpenPack dataset provides a wealth of data for classifying actions in industrial settings, collected from multiple sensors in a controlled environment, and proposes a competition to develop a system capable of recognizing the actions performed by an operator. This work aimed to employ a series of multi-modal combined models that process data from inertial sensors and visible cameras to classify each action performed by the operator, achieving results comparable to those of the top competitors who participated in the competition proposed by the dataset creators. To achieve this, we leveraged the features of convolutional neural networks and recurrent neural networks for data from inertial sensors, and graph-based neural networks for the visual stage through keypoint-based pose estimation.
Descripción
Categorías UNESCO
Palabras clave
reconocimiento de acciones humanas, dataset, deep learning, aprendizaje automático, redes neuronales convolucionales, redes neuronales recurrentes, redes neuronales convolucionales basadas en grafos, keypoints, estimación de pose, sensores inerciales, LiDAR, imagen en profundidad, human action recognition, machine learning, convolutional neural networks, recurrent neural networks, graph convolutional networks, pose estimation, inertial sensors, depth image
Citación
Centro
Facultades y escuelas::E.T.S. de Ingeniería Informática
Departamento
Inteligencia Artificial