Semantic Segmentation-Based Lane-Level Localization Using Around View Monitoring System

Abstract

Lane-level localization is a fundamental task for autonomous driving. As front cameras are easily disturbed by dynamic objects in urban environments, this paper presents an accurate lane-level localization approach using the around view monitoring (AVM) system. The paper proposes to detect the road features (i.e., road boundaries and road markings) based on pixel-wise semantic segmentation of raw fisheye images. The method can detect various types of road features and exclude dynamic objects from the localization. To address the problem of AVM-based localization with road features of different characteristics, this paper proposes coarse-scale localization (CSL) and fine-scale localization (FSL) methods for high-accurate localization. The CSL method leverages the road boundaries to provide an initial position; the FSL method estimates a high-accuracy position by matching nearby road markings with the map. The experiments in urban environments demonstrate that the proposed approach achieves centimeter-level localization accuracy with five centimeters in the lateral direction and seventeen centimeters in the longitudinal direction.