Tools for pre-processing and dimensional reduction of images in the recognition and deflection of obstacles by a mobile robot

• Grant number: 22/02382-7
• Support Opportunities: Scholarships in Brazil - Scientific Initiation
• Start date: May 01, 2022
• End date: April 30, 2023
• Field of knowledge: Physical Sciences and Mathematics - Computer Science - Computing Methodologies and Techniques
• Principal Investigator: Glauco Augusto de Paula Caurin
• Grantee: Luís Eduardo de Souza Cintra
• Host Institution: Escola de Engenharia de São Carlos (EESC). Universidade de São Paulo (USP). São Carlos , SP, Brazil

Abstract

The human brain, in addition to the simple interpretation of electrical signals, requires that several brain regions act together to synthesize learning resulting from sensory experiences, an essential skill for living beings. Transmitting this remarkable biological capacity to machines and digital systems has become the goal of Artificial Intelligence, a high-level computational invention that aims to reproduce the brain action of learning and apprehension on a variety of devices. From this large area, other topics were derived that, together, make up an overview of numerous tools for improving Machine Learning and Artificial Neural Networks, such as Deep Learning, Reinforcement Learning and Deep Reinforcement Learning. The AI training and implementation models cover several areas, such as the financial market, medicine and robotics; however, the implementation of this technology can become computationally expensive, especially when applied to large databases with millions of features, which requires enormous processing power that, when limited, significantly increases the training time. Thus, this work seeks to study the Dimensionality Reduction methods, which consists of data manipulation techniques so that the dimensions of its representation can be reduced and, therefore, more easily interpreted and learned by the machine, offering a minimum informational loss. For this, the development environment will be based on a mobile robot that, having a camera for capturing images and visualizing obstacles, will transmit each frame to a processing network responsible for applying the dimensional reduction and, consequently, improving the control system performance of the robotic device. This study also aims to compare different Dimensionality Reduction techniques, divided between linear (e.g. PCA, LDA) and nonlinear (e.g. Isomap, LLE, Kernel-PCA), enabling to find the most efficient one, thus allowing changes to be studied to improve the relationship between image quality and training time.(AU)