Visual urban road features detection using Convolutional Neural Network with application on vehicle localization

• Autor(a) principal: Horita, Luiz Ricardo Takeshi
• Data de Publicação: 2018
• Tipo de documento: Dissertação
• Idioma: eng
• Título da fonte: Biblioteca Digital de Teses e Dissertações da USP
• Texto Completo: http://www.teses.usp.br/teses/disponiveis/18/18153/tde-10122018-152247/
• topic: Aprendizado de máquina; Convolutional Neural Network; Curb detection; Detecção de guia; Detecção de sinalização horizontal; Localização de veículos; Machine learning; Rede Neural Convolutiva; Road marking detection; Stereo vision; Vehicle localization; Visão estéreo

Curbs and road markings were designed to provide a visual low-level spatial perception of road environments. In this sense, a perception system capable of detecting those road features is of utmost importance for an autonomous vehicle. In vision-based approaches, few works have been developed for curb detection, and most of the advances on road marking detection have aimed lane markings only. Therefore, to detect all these road features, multiple algorithms running simultaneously would be necessary. Alternatively, as the main contribution of this work, it was proposed to employ an architecture of Fully Convolutional Neural Network (FCNN), denominated as 3CSeg-Multinet, to detect curbs and road markings in a single inference. Since there was no labeled dataset available for training and validation, a new one was generated with Brazilian urban scenes, and they were manually labeled. By visually analyzing experimental results, the proposed approach has shown to be effective and robust against most of the clutter present on images, running at around 10 fps in a Graphics Processing Unit (GPU). Moreover, with the intention of granting spatial perception, stereo vision techniques were used to project the detected road features in a point cloud. Finally, as a way to validate the applicability of the proposed perception system on a vehicle, it was also introduced a vision-based metric localization model for the urban scenario. In an experiment, compared to the ground truth, this localization method has revealed consistency on its pose estimations in a map generated by LIDAR.