Uso de machine learning para medir productividad a cuadrillas de trabajo usando videos y detección de imágenes a través de cartas de balance
Cargando...
Archivos
Fecha
2022
Profesor/a Guía
Facultad/escuela
Idioma
es
Título de la revista
ISSN de la revista
Título del volumen
Editor
Universidad Andrés Bello
Nombre de Curso
Licencia CC
Licencia CC
Resumen
Expertos y líderes en la industria de la construcción señalan que nuevas tecnologías basadas en
algoritmos de inteligencia artificial son una de las prioridades y uno de los principales factores para
aumentar la productividad. En Chile el crecimiento anual de la productividad durante las dos
últimas décadas en el sector de la construcción no ha experimentado variación. Debido a esto se
propone medir la productividad, a través de un algoritmo de Machine Learning, dado que los
resultados adquiridos entregan información confiable y veraz aportando mejoras en los procesos
constructivos y a la toma rápida de decisiones.
El proyecto de investigación se basa en el reconocimiento de posturas corporales de trabajadores
de una cuadrilla de trabajo utilizando videos e imágenes con Machine Learning. El algoritmo
YOLO se basa en técnicas de aprendizaje automático capaz de reconocer posturas corporales de
los trabajadores, estas posturas se asocian a Trabajos Productivos y Contributorios, definidos
previamente en una taxonomía que describe las posturas corporales.
El algoritmo YOLO se entrenó con un total de 1500 imágenes extraídas de 74 videos grabados en
360°, con un total de 5 horas de grabación a una cuadrilla de trabajadores involucrados en la
actividad de instalación de moldaje de losas en la etapa de obra gruesa de un edificio habitacional
de 16 pisos, ubicado en Av. Vicuña Mackenna, comuna de San Joaquín. El algoritmo presentó
resultados experimentales de una precisión media (mAP 0.5) mayor al 85% para cada una de las
tareas asociadas a la actividad.
Para la etapa de medición de productividad se utilizaron grabaciones realizadas con una cámara
fija capturando la actividad completa, de estos videos se extrajeron imágenes cada 5 segundos. A
partir del algoritmo YOLO entrenado se obtuvo como resultado las detecciones de las posturas
corporales de los trabajadores para cada una de las tareas: instalación de encintado perimetral (IEP),
instalación de puntales aplomados (IPA), instalación de vigas (IV) e instalación de placas (IP). Esta
última tarea se describe en detalle durante el desarrollo de la investigación, ya que es la que obtuvo
mejores resultados en cuanto al entrenamiento y diagnóstico.
Los resultados obtenidos por el algoritmo YOLO se presentaron en Cartas de Balance, herramienta
que permite generar diagnósticos productivos de trabajo, identificando los tiempos de trabajos
Productivos y Contributorios. Estas Cartas de Balance se compararon con las Cartas de Balance
realizadas de forma manual, con el fin de saber qué tan bueno es el algoritmo para reconocer estas
posturas corporales. Para la tarea de instalación de placas (IP) se obtuvo como resultado un 72%
de imágenes detectadas por el algoritmo para el Trabajo Productivo, un 0% de imágenes detectadas
para el Trabajo Contributorio y un 58% para el total de imágenes detectadas por el algoritmo.
Finalmente, la metodología propuesta se validó, a través de una encuesta realizada a 3 profesionales
de terreno y 2 académicos que se desempeñan en el sector de la construcción. La encuesta promedió
un total de 3.2 en la escala de Likert de 1 a 4. Los comentarios y observaciones se tomaron en
cuenta para finalmente realizar la propuesta definitiva de la metodología de investigación.
Experts and leaders in the construction industry point out that new technologies based on artificial intelligence algorithms are one of the priorities, and one of the main factors to increase productivity. The annual productivity growth in the Chilean construction industry during the last two decades has not changed. Therefore, we propose to measure productivity using a Machine Learning algorithm that delivers reliable productivity information. This could help provide improvements in construction processes, and it could facilitate decision-making. This research project is based on recognition of workers' body postures from a work crew using videos and images with Machine Learning. The YOLO algorithm is based on Machine Learning techniques capable of recognizing body postures of workers. These postures are associated with Productive and Contributory work, previously defined in a taxonomy that describes such body postures. The YOLO algorithm was tested on 1500 images extracted from 74 videos recorded in 360 degrees using a GoPro action camera. We recorded over 5 hours of crew work doing the slab formwork installation process. Our experimental results have a mean average precision (mAP 0.5) greater than 85% for each of the tasks associated with the activity. For the productivity measurement stage, recordings made with a fixed camera were used. We captured the complete activity and we extracted still images every 5 seconds. From the trained YOLO algorithm, the detections of the body postures of the workers for each of the tasks were obtained as a result: perimeter formwork installation (IEP), installation of support struts (IPA), installation of beams (IV) and installation of formwork boards (IP).This last task is described in detail during this research, since it showed the best results in terms of training and productivity diagnosis. The results obtained by the YOLO algorithm were presented in Balance Charts, a tool that allows generating productive work diagnoses. It identifies time spent on Productive and Contributory work. These Balance Charts were compared with Balance Charts built manually, in order to know how good the algorithm is to recognize these body postures. For the installation of formwork boards (IP), We obtained a 72% of images detected by the algorithm for Productive Work, while 0% of images were detected for Contributory Work. A total of 58% of detections were achieved by the algorithm. Finally, the proposed methodology was validated through a survey applied to 3 field professionals and 2 academic experts from the construction industry. The survey averaged a total of 3.2 on the Likert scale from 1 to 4. Their comments and observations were considered for the final proposal of the research methodology.
Experts and leaders in the construction industry point out that new technologies based on artificial intelligence algorithms are one of the priorities, and one of the main factors to increase productivity. The annual productivity growth in the Chilean construction industry during the last two decades has not changed. Therefore, we propose to measure productivity using a Machine Learning algorithm that delivers reliable productivity information. This could help provide improvements in construction processes, and it could facilitate decision-making. This research project is based on recognition of workers' body postures from a work crew using videos and images with Machine Learning. The YOLO algorithm is based on Machine Learning techniques capable of recognizing body postures of workers. These postures are associated with Productive and Contributory work, previously defined in a taxonomy that describes such body postures. The YOLO algorithm was tested on 1500 images extracted from 74 videos recorded in 360 degrees using a GoPro action camera. We recorded over 5 hours of crew work doing the slab formwork installation process. Our experimental results have a mean average precision (mAP 0.5) greater than 85% for each of the tasks associated with the activity. For the productivity measurement stage, recordings made with a fixed camera were used. We captured the complete activity and we extracted still images every 5 seconds. From the trained YOLO algorithm, the detections of the body postures of the workers for each of the tasks were obtained as a result: perimeter formwork installation (IEP), installation of support struts (IPA), installation of beams (IV) and installation of formwork boards (IP).This last task is described in detail during this research, since it showed the best results in terms of training and productivity diagnosis. The results obtained by the YOLO algorithm were presented in Balance Charts, a tool that allows generating productive work diagnoses. It identifies time spent on Productive and Contributory work. These Balance Charts were compared with Balance Charts built manually, in order to know how good the algorithm is to recognize these body postures. For the installation of formwork boards (IP), We obtained a 72% of images detected by the algorithm for Productive Work, while 0% of images were detected for Contributory Work. A total of 58% of detections were achieved by the algorithm. Finally, the proposed methodology was validated through a survey applied to 3 field professionals and 2 academic experts from the construction industry. The survey averaged a total of 3.2 on the Likert scale from 1 to 4. Their comments and observations were considered for the final proposal of the research methodology.
Notas
Memoria (Ingeniero Civil)
Palabras clave
Aprendizaje de Máquina, Industria de la Construcción, Productividad, Postura Humana, Innovaciones Tecnológicas, Investigaciones