Tony Ferreira

Abstract
The robotization and automation of tasks are relevant processes and of great relevance to be considered nowadays. This work aims to turn the manual action of assigning the score for the robotic competition Robot at Factory 4.0 by an automatic referee. Specifically, the aim is to represent the real space in a set of computational information using computer vision, localization and mapping techniques. One of the crucial processes to achieve this goal involved the adaptive calibration of the parameters of a digital camera through visual references and tracking of objects, which resulted in a fully functional, robust and dynamic system that is capable of mapping the competition's objects accurately and correctly performing the referee's tasks.

Abstract
Quality control inspection systems are crucial and a key factor in maintaining and ensuring the integrity of any product. The quality inspection task is a repetitive task, when performed by operators only, it can be slow and susceptible to failures due to the lack of attention and fatigue. This work focuses on the inspection of parts made of high-pressure diecast aluminum for components of the automotive industry. In the present case study, last year, 18240 parts needed to be reinspected, requiring approximately 96 hours, a time that could be spent on other tasks. This article performs a comparison of four deep learning models: Faster R-CNN, RetinaNet, YOLOv7, and YOLOv7-tiny, to find out which one is more suited to perform the quality inspection task of detecting metal filings on casting aluminum parts. As for this use-case the prototype must be highly intolerant to False Negatives, that is, the part being defective and passing undetected, Faster R-CNN was considered the bestperforming model based on a Recall value of 96.00%.