Real-time image absolute orientation framework with Optical-GNSS data fusion for lightweight UAVs

For lightweight unmanned aerial vehicles (UAVs), real-time absolute camera orientation is critical for photogrammetric mapping. However, achieving absolute accuracy under inertial-inaccessible constraints remains challenging due to the scale drift of visual odometry and the non-stationary noise of global positioning measurements. To address these operational challenges, this paper presents a practical hierarchical visual–GNSS fusion framework designed for real-time onboard deployment. The system adopts a decoupled multi-threaded architecture that explicitly separates local state estimation from global optimization to ensure real-time performance on resource-limited platforms. This design avoids tightly coupled optimization schemes in which uncertainty propagation is entangled with nonlinear estimation. Within this framework, an asynchronous filtering module evaluates GNSS statistical consistency using normalized innovation statistics and generates reliability-aware confidence signals. These signals are propagated across estimation layers to modulate the GNSS information matrix in the global factor graph, while soft pose priors provide complementary regularization during intervals of degraded observations. To maintain frame-rate throughput under resource-limited onboard conditions, real-time metric tracking and sliding-window alignment are executed causally on the main thread, while computationally intensive global pose-graph optimization and bundle adjustment are offloaded to an asynchronous background thread for delayed drift refinement. Evaluations on selected public benchmarks and self-collected datasets indicate that the proposed framework yields sub-meter absolute camera positioning accuracy (0.25–0.77 m) under low-overlap, repetitive texture, and non-stationary GNSS degradation. Operating on a lightweight CPU-only platform, the complete system maintains a processing throughput of over 3 FPS for 20-megapixel photogrammetric imagery and 28 FPS for HD video streams, indicating its practical onboard feasibility.