Generating road grid maps and path plans for self-driving cars using laser remission data and deep neural networks

• Autor(a) principal: Carneiro, Raphael Vivacqua
• Data de Publicação: 2024
• Tipo de documento: Tese
• Idioma: por
• Título da fonte: Repositório Institucional da Universidade Federal do Espírito Santo (riUfes)
• Texto Completo: http://repositorio.ufes.br/handle/10/18224
• topic: Carros autônomos; Remissão de laser; Mapas de grade; Ciência da Computação

This work proposes the use of deep neural networks (DNN) for solving the problem of inferring the location of drivable lanes of roadways and their relevant properties such as the lane change right-of-way, even if the line markings are poor or absent. This problem is relevant to the operation of self-driving cars which requires precise maps and precise path plans. Our approach to the problem is the use of a DNN for semantic segmentation of LiDAR remission grid maps into road grid maps. Both LiDAR remission grid maps and road grid maps are square matrices in which each cell represents features of a small 2D-squared region of the real world (e.g., 20cm × 20cm). A LiDAR remission grid map cell contains the information about the average intensity of laser reflection remission on the surface of that particular place. A road grid map cell contains the semantic information about whether it belongs to either a drivable lane or a line marking or a non-drivable area. The semantic codes associated with the road map cells contain all information required for building a network of valid paths, which are required for self-driving cars to build their path plans. Our proposal is a novel technique for the automatic building of viable path plans for self-driving cars. In our experiments we use the self-driving car of UFES, IARA (Intelligent Autonomous Robotic Automobile). We built datasets of manually marked road lanes and use them to train and validate the DNNs used for the semantic segmentation and the generation of road grid maps from laser remission grid maps. The results achieved an average segmentation accuracy of 94.7% in cases of interest. The path plans automatically generated from the inferred road grid maps were tested in the real world using IARA and has shown performance equivalent to that of manually generated path plans.