Detecção vsual de caminhos através de aprendizado profundo para a navegação de robôs móveis

Abstract

Detecting the path that a vehicle should follow is an essential task in autonomous robot navigation. In recent years, deep learning-based approaches have outperformed traditional computer vision techniques in path detection, due to the ability of neural networks to extract features directly from data, making detection more robust and accurate across different scenarios. This work presents a transfer learning-based approach to tackle the problem of visual path detection for mobile robot navigation, using a factorized residual convolutional neural network. The work presents a comparison between the proposed model and other convolutional neural network architectures, as well as a comparative case study with a system based on deterministic computer vision. Experimental results indicate the promise of the proposed model, which is capable of detecting paths even in the presence of discontinuities and variations in lighting, achieving an F1 score of 0.9166 and an average inference time of 6.15 ms. The proposed architecture also achieves a good balance between accuracy and efficiency, making the system suitable for mobile robotics applications.