PO-GVIO: A GPS-aided Visual-Inertial Odometry with Pose-only Representation

With the advancement of autonomous navigation, multi-sensor fused localization systems have attracted significant research interest. This paper proposes PO-GVIO, a GPS-aided visual-inertial odometry (VIO) system based on the Multi-State Constraint Kalman Filter (MSCKF). The VIO component adopts a pose-only representation, which eliminates explicit reconstruction of 3D feature points and instead constructs observation equations directly from camera poses and visual measurements. This approach effectively reduces linearization errors and avoids update latency. Nevertheless, the increased sensitivity to observation noise, especially due to dynamic objects and feature mismatches, may adversely affect the system accuracy. To address this issue, a self-adaptive outlier rejection strategy is introduced to improve system robustness. Additionally, a tiered fusion strategy is proposed to enhance performance in GPS-degraded environments. Experimental results demonstrate that the pose-only representation significantly outperforms conventional methods in localization accuracy, and the outlier rejection strategy notably enhances system stability. With GPS integration, PO-GVIO achieves comparable performance to IC-GVINS on the public datasets. Furthermore, the system's reliability is validated on private dataset, demonstrating accurate and smooth localization in real-world outdoor scenarios.