Hybrid Map for Autonomous Commercial Vehicles - Global localization using topological mapping and machine learning

Abstract

We propose and investigated a novel method for global navigation and localization of an autonomous commercial vehicle within a con ned area using a hybrid map. The hybrid map is based on a topology using nodes and edges where signi cant places are adapted as nodes. The hybrid map is able to store di erent type of machine learning algorithms and its exible design allows the topology to be easily extended. The hybrid map operates using a node detector algorithm complimented with a node classi cation algorithm for increased robustness. The machine learning algorithms uses two dimensional lidar data as inputs exclusively. When it comes to the detection of nodes, performance evaluation showed that the Adam method are superior to the common gradient descent method when training feed forward neural networks in the considered scenario. In order to classify the nodes, one class support vector machines are preferred. The performance of the hybrid map system was further on evaluated by implementing it on a Raspberry Pi 3 to prove its simplicity. In conclusion, our results suggest that the system has potential for implementation in a real vehicle. However, it needs further veri cation and improvements to ensure a robust system and for it to be useful as a real application.