Visual-Inertial-GNSS Fusion Positioning for Vehicles With Deep-Learning-Based Feature Extraction and Outlier Detection

Abstract

Reliable and continuous high-precision positioning is a fundamental requirement for navigation, guidance, and control of intelligent mobile platforms, such as autonomous vehicles, drones, and robots. However, achieving such positioning in complex urban environments remains a significant challenge due to GNSS signal obstructions and multipath effects. Moreover, traditional visual feature extraction algorithms struggle with variations in lighting, perspectives, and noise, limiting their adaptability. Consequently, relying solely on a single sensor often fails to provide a stable and accurate positioning solution. This paper addresses the challenges encountered in typical urban environments and proposes a Visual-Inertial-GNSS fusion positioning framework, in which deep learning-based feature extraction is applied to the visual front-end process. The deep learning-based feature extraction approach fully exploits image information to derive more accurate and robust features, thereby enabling reliable feature matching. Additionally, we introduce a K-means clustering-based outlier detection method to remove dynamic features and mismatched features, thereby enhancing optical flow tracking and overall positioning accuracy. Experimental results in urban environments demonstrate that the proposed approach achieves high-precision positioning results under challenging GNSS conditions and complex visual environments, with improvements of (29.6%, 10.0%, 7.3%) over SIFT and (47.8%, 7.8%, 15.5%) over ORB in the North-East-Down (N-E-D) directions. Moreover, in dynamic road environments, the proposed outlier detection method achieves accuracy improvements of 19.6%, 3.0%, and 4.0% in the N-E-D directions compared to the RANSAC-based method.