Junction Points Aided Point-Line-Based GNSS/INS/Vision Integration Outlier Exclusion and Localization in Degradation Environments

The development of autonomous driving leads to a higher requirement for high-precision positioning in complex urban areas. However, the inherent limitations of single sensors may cause huge positioning errors in complex urban areas. To mitigate this issue, multiple sensors are fused based on their complementarity. Among them, the GNSS/INS/vision fusion framework has been widely deployed and mainly focuses on corner points and line features for the outdoors. However, the degradation of corner points and the unstable endpoints of line features make it difficult to achieve a high-precision outlier rejection and positioning. To address this issue, we present a point-line-based GNSS/INS/vision fusion system with the aid of junction points for outlier exclusion and location. An RANSAC method used for outlier rejection is proposed, where the essential matrix is optimized by both corner and junction points to improve the outlier rejection performance in low-texture environments. Through the association by line-junction-line descriptors, the inliers of junction points are used to reject the outliers of line features. For state estimation, more accurate corner points and line features combined with additional junction points are employed to improve GNSS/INS integration accuracy under degraded GNSS signal conditions. The proposed GNSS/INS/Vision integration system can significantly improve the positioning performance when GNSS and corner points are degraded, such as trees and buildings. Two urban infield experiments validate the proposed system’s efficacy, reducing maximum horizontal positioning errors by 20.93% and 39.74% compared with corner-line-based integration systems. Additional evaluations on public datasets demonstrate our system’s superiority, achieving greater accuracy than the state-of-the-art GNSS/INS/vision fusion framework VINS-Fusion and IC-GVINS. The positioning errors are reduced by 75.90% and 54.61%.