Detection of vehicles and buildings in drone aerial images

• Autor(a) principal: Amante, Rita Filipa dos Santos
• Data de Publicação: 2022
• Tipo de documento: Dissertação
• Idioma: eng
• Título da fonte: Repositório Científico de Acesso Aberto de Portugal (Repositórios Cientìficos)
• Texto Completo: http://hdl.handle.net/10773/34992
• topic: Artificial intelligence; Machine learning; Deep learning; Transfer learning; Computer vision; Object detection; UAV

The need to develop software for aerial image analysis, captured by Unmanned Aerial Vehicles, has increased over the years because their use has become more prevalent in different day-to-day scenarios. Object detection, a Computer Vision technique, is one of the most explored problems in this area and consists of identifying and locating objects in images or videos, with the help of Artificial Intelligence technologies. The aim of this dissertation is to analyze the performance of Deep Learning algorithms for detecting vehicles and buildings in aerial images. Two of the main algorithms described in literature, Faster R-CNN and YOLO, the latter in the third and fifth versions, were chosen to verify which one is capable of better performance. The dataset provided by the Portuguese Military Academy, which was annotated and pre-processed, was used for the training of each algorithm and the performance of tests. The results obtained in the abovementioned dataset demonstrate that there is a considerable discrepancy between the two algorithms, both in terms of performance and speed. Faster R-CNN only proved to be superior to the two versions of YOLO in terms of training speed, as it was the algorithm that required less time for training. Among the versions of YOLO, the fifth version showed the best results.