Desarrollo de un sistema de visión artificial multiagente para la segmentación de vidrio en ventanas de rascacielos.

Abstract

The number of skyscrapers under construction and built in recent years demonstrates the growth dynamics of skyscrapers and, with the requirements and technologies used to clean their windows, show a clear need to develop new alternatives that perform this task in a safer, faster and more accurate way. This work presents the development of a multi agent artificial vision system for glass segmentation in skyscraper windows, which is designed to be attached in the future into a robotic arm prototype and a drone, in order to perform the work of skyscraper window cleaning. The function of the developed system is to segment the glass and, from the segmentation, obtain the parameters of the height, width and position in the X, Y and Z axes of the glass's center in relation to the camera's position. The research presents a theoretical introduction of the methods used for the development of the mentioned system, a state of the art of artificial vision, and the methodology and results that guided the selection of the methods used. The results include the compilation of a database of skyscraper windows with a wide variety of physical and lighting conditions; a preliminary glass segmentation program using edge detection algorithms such as Canny, Sobel, Prewitt, Roberts and Laplacian of the Gaussian, region properties focused on rectangular areas, and line detection using Hough transform. Additionally, the implementation of artificial intelligence techniques such as convolutional neural networks is presented along with semantic segmentation, and all the methods set forth above integrated into a multi-agent artificial vision system for glass segmentation using a decision tree. Lastly, a way to achieve the conversion of pixels into the metric system of the parameters delivered by the multi-agent system is proposed, in order to demonstrate that the developed system can be implemented. The above is achieved by following the mathematical equations of the relationship between the image taken by the camera, the measurement made by a distance sensor and the intrinsic parameters of the camera.