A visual SLAM navigation method for unmanned aerial vehicles based on sky polarized light assistance
Abstract
Visual SLAM (Simultaneous Localization and Mapping) is a core technology for UAV autonomous navigation. However, it suffers from cumulative errors and heading drift in texture-less environments (e.g., sky, water surfaces) or during longterm operation. Traditional methods relying on GPS or inertial sensors face limitations in denied environments (e.g., indoors, canyons). This study proposes a visual SLAM framework enhanced by sky polarized light information. A polarization camera captures atmospheric scattering patterns to estimate the solar azimuth as an absolute directional constraint, fused with visual features and IMU data. Key innovations include: Polarized Light-Vision Tightly-Coupled Model: Integrates polarization angle observations into SLAM back end optimization to suppress heading drift. Dynamic Weather-Adaptive Algorithm: Adjusts polarization weighting based on cloud detection for robustness under complex meteorological conditions. Lightweight Embedded Deployment: Optimizes computational efficiency for UAV onboard platforms. Simulations and real-world tests show a 62% reduction in position error and 75% improvement in heading accuracy compared to traditional visual SLAM. Stable navigation exceeding 30 minutes was achieved in GPS-denied environments. Sky polarized light assistance significantly enhances the long-term reliability of visual SLAM, offering a new paradigm for UAV autonomous navigation in signal-denied scenarios. This work holds great value for bio-mimetic navigation and extreme environment operations.