Segmentación de imágenes urbanas de un dron mediante Deep Learning

Abstract

El campo de la Inteligencia Artificial está adquiriendo una relevancia considerable en diversos ámbitos y sectores de la sociedad. Uno de ellos donde la Inteligencia Artificial ha destacado significativamente, es en la clasificación de imágenes y detección de objetos dentro de las mismas. Gracias a estos avances se ha conseguido desarrollar aplicaciones tan dispares como: conducción autónoma, diagnostico médico o automatización industrial. Este trabajo explora el área de la Visión por Computador para la detección de zonas de aterrizaje seguras para drones empleando, para ello, la segmentación de imágenes urbanas. Uno de los propósitos de este trabajo es presentar, tanto el marco teórico, como práctico, para desarrollar este modelo de segmentación de imágenes. Para este fin, se emplean técnicas de Deep Learning (Aprendizaje Profundo), en particular, Redes Neuronales Convolucionales, bajo la arquitectura conocida como UNET. A pesar de que los resultados obtenidos alcanzan una alta precisión, considero adecuado mencionar, que no se puede asumir como una solución única y definitiva. Finalmente, el trabajo se complementa con una aplicación web, que permite utilizar el sistema de detección de zonas de aterrizaje de forma eficiente y sencilla. Con ello, una persona puede supervisar y analizar los resultados obtenidos por la aplicación y proporcionar una cierta explicación de los mismos.

Today, Artificial Intelligence (AI) applications are growing up day by day reaching almost everything in our life. In particular, it becomes more popular in Artificial Vision problems as image classification or semantic segmentation, where the Deep Learning has been widely used. Consequently, several intelligence tasks as autonomous driving, medical diagnosis or industrial automation have experimented a very important successes. This work is focused on drone safe landing zones detection in a satellite (or done) photo. It can be considered as one more image segmentation example, if the corresponding ground truth is available as it taken place in practice. To address that problem, a certain Artificial Vision algorithm will be used. More precisely, a Convolutional Neural Network under a UNET architecture has been selected. First of all, a theoretical study for understanding that AI model is introduced. After that, an experimental framework will be chosen, explained and implemented. The next step will show the results obtained. In spite they reach a high accuracy level, it doesn’t imply that system proposed will be considered as definitive, since another models should be experimented to discard them or to taken into consideration. Unfortunately, that task can not be affordable in this work. Finally, an application web based has been made to manage the AI model trained. Once the photo was entered, this program outputs the drone safety landing zones estimated by the system, as well as four more kinds of zones depicted in this text. As conclusion, this project has showed one more time again, that UNET architecture could be used as the basic algorithm to image semantic segmentation offering a very good results.