ODUFT - dataset RGB-D para aplicações de detecção e reconhecimento de objetos domésticos na RoboCup@Home Brasil

Abstract

Service robotics is emerging as a growing field, significantly impacting daily life. Service robots, autonomous devices designed for daily tasks, play a vital role, particularly in assisting people with disabilities. Competitions, such as RoboCup@Home, drive advancements in service robotics by promoting the optimization of existing solutions and the creation of innovations. Teams, focused on enhancing service robot platforms, face complex challenges such as navigating dynamic environments, detecting and recognizing people and objects, speech recognition, natural language processing, and object manipulation in unstructured domestic environments. Highlighting the interaction between activities in the competition, object detection and recognition emerge as crucial elements. Most teams turn to convolutional neural networks (CNNs) to identify the presence and location of objects, justifying this choice due to the rich representation of visual information, the effectiveness of architectures, and the availability of comprehensive datasets. CNNs are designed to efficiently process image data, learning fundamental hierarchical features. Creating image datasets is a challenging task that demands time and resources. Utilizing automated approaches, whether through internet image search engines or vision systems, images are collected and stored in defined locations for retrieval. Given the lack of massive datasets for the RoboCup@Home competition, this thesis proposal aims to create and provide a set of unprocessed images, in raw format, for teams participating in the @Home League of RoboCup Brazil. This document delineates the rationale and methodology developed for the creation and dissemination of a dataset comprising unprocessed and fully annotated images in the YOLO format. Acknowledging the non-uniform presentation of objects in real-world environments, synthetic images were generated to characterize elements such as partial occlusion, variations in illumination, and the presence of noise, subsequently integrated into the dataset. Upon completion of these preparatory stages, the study proceeded to the data validation phase, conducted by executing a benchmark against other publicly available datasets, which involved conducting a comparative benchmark against other publicly available datasets. This initiative seeks to contribute significantly to the advancement of service robotics, providing valuable resources for the community involved in RoboCup@Home Brazil.