PO-P3VINS: Lightweight PPP-Visual-Inertial Tightly-Coupled Positioning System With Pose-Only Visual Constraint

Abstract

Precise Point Positioning (PPP) enables global positioning without base stations but suffers from prolonged convergence time, limiting its real-time applicability in autonomous systems. Meanwhile, Visual-Inertial Navigation System (VINS) provides smooth short-term motion estimates but accumulates unbounded drift over long durations. PPP and VINS are therefore complementary, motivating their tight integration. However, conventional VINS model introduces substantial computational complexity due to explicit landmark state estimation. To improve computational efficiency, we propose PO-P3VINS, a lightweight tightly-coupled positioning system that integrates PPP and VINS through pose-only (PO) visual constraint. By reformulating visual constraints without maintaining explicit landmark depth parameters in the state vector, the proposed method effectively reduces the optimization state dimension and computational complexity. PO-P3VINS is evaluated on both public and self-collected datasets. Experimental results demonstrate that PO-P3VINS achieves a $1.6\times $ – $2.3\times $ computational speedup, corresponding to approximately 38%–56% reduction in the cumulative optimization time, compared with conventional bundle adjustment-based PPP-visual-inertial methods, while maintaining comparable positioning accuracy. Note to Practitioners—This paper is motivated by the need for reliable and efficient positioning in autonomous systems that operate for long periods in urban environments, such as intelligent vehicles, surveying devices, and mobile robots. In these applications, satellite positioning can provide global location information, but its accuracy may degrade or require a long time to stabilize. Cameras and inertial sensors can provide smooth short-term motion estimates, but they may drift over time and can increase the computational load of the system. This paper presents a lightweight positioning approach that combines satellite, camera, and inertial measurements while reducing the amount of visual information that must be estimated during optimization. Instead of explicitly estimating many visual feature depths, the proposed method uses pose-only visual constraints. This design lowers the computational burden while maintaining positioning accuracy. In practical terms, the method can help embedded or handheld platforms achieve more stable real-time operation, reduce processor load, and improve deployability in resource-constrained autonomous systems. The results show that the proposed system can reduce optimization time while preserving comparable or improved positioning accuracy in both public and self-collected urban datasets. However, the current implementation uses float ambiguity estimation and has not yet incorporated ambiguity resolution, which may further improve accuracy after convergence. Future work will investigate this extension and evaluate the system in broader operating conditions. The proposed idea may also be useful for other multi-sensor automation systems requiring long-duration, real-time localization.