Novel techniques for mapping and localization of self driving cars using grid maps
Nenhuma Miniatura disponível
Data
2019-09-02
Autores
Mutz, Filipe Wall
Título da Revista
ISSN da Revista
Título de Volume
Editor
Universidade Federal do Espírito Santo
Resumo
This work proposes novel techniques for building grid maps of large-scale environments, and for estimating the localization of self-driving cars in these maps. The mapping technique is employed for creating occupancy, reflectivity, colour, and semantic grid maps. The localization is based on particle filters. New methods for computing the particles’ weights using semantic and colour information are presented. The deep neural network DeepLabv3+ is used for visual semantic segmentation of images captured by a camera. The estimation of the vehicle poses for mapping is modelled as a Simultaneous Localization and Mapping (SLAM) problem. The values of the poses are obtained by using the GraphSLAM algorithm to fuse odometry and GPS data. These values are refined using loop-closure information. The optimized poses are used for building maps of the environment. The self-driving cars localizations are computed in relation to these maps. The mapping and localization techniques were evaluated in several complex and large-scale environments using a real self-driving car – the Intelligent and Autonomous Robotic Automobile (IARA). The impact of using different types of grid maps in the localization accuracy as well as its robustness to adverse conditions of operation (e.g., variable illumination, and intense traffic of vehicles and pedestrians) were evaluated quantitatively. As far as we know, the mapping and localization techniques, the methodology for producing the localization ground truth, and the evaluation of which type of grid map leads to more accurate localization are novelties
Descrição
Palavras-chave
Palavra-chave , Carros autônomos , Localização , Mapas de grade , Robotics , Self-driving cars , Localization , Grid maps