Reconhecimento de objetos em imagens RGB-D visando estimativa da pose em tempo real

Abstract

Computer Vision is a field of study in Artificial Intelligence that seeks to employ Machine Learning algorithms to provide applications with the ability to detect visual patterns, classify, and reconstruct objects in multiple dimensions. In this context, this work proposes a substantial improvement based on the conclusions of another study, previously developed. This improvement consists of adding a stage to the execution flow implemented in the original work, which will focus on the detection and classification of objects in two-dimensional scenes, thus seeking to improve the overall result and performance. For this purpose, YOLO, an object detection model, was used in its YOLOv4, YOLOv4-tiny, YOLOv7, and YOLOv7-tiny versions. Thus, it was possible to apply the proposed methodology and perform training and testing of the models using the RGB Scenes dataset. The results proved the effectiveness of YOLO with a mAP value of 99.84% in the best scenario. Furthermore, the processing speed achieved in frames per second was more than ten times higher than that originally achieved, reaching 330 FPS. This demonstrated that the initial hypothesis was correct, given the significant improvement in the performance of the detection task.