Ground mobile robot localization algorithm based on semantic information from the urban environment

Abstract

This paper presents the SLAM algorithm, which use the semantic information extracted from the urban environment to increase the accuracy of ego-vehicle localization in ORB-SLAM2 system. For this purpose, a semantic segmentation module is added to the standard algorithm to assign an object on each frame to one of a given set of classes. The CARLA Simulator was used as a simulation environment, which generates a photorealistic urban environment with the ability to run an arbitrary number of active elements in it, which usually make localization difficult, causing interference with the system. Based on the environment, a training dataset for semantic segmentation was collected. The training dataset consists of 3,696 pairs of city images and corresponding segmentation masks in which each pixel corresponds to one of 23 semantic labels. Using this dataset, the DeepLabV3+ segmentation model was trained with mean per-class IoU metric equals to 81.48%. By using semantic information to filter potentially dynamic objects and matching key points, we were able to increase the localization accuracy relative to the base algorithm by an average of 23% and build a semantic map of the environment.