Adaptive UAV Navigation Fusion with Scene Matching and Mixed Credibility UKF

To address the degradation of UAV positioning accuracy caused by variations in sensor performance under changing environments, a scene-matching-based multi-sensor fusion algorithm, termed mixed credibility unscented Kalman filter (MCUKF), is proposed. The method jointly considers environmental and motion scenes for adaptive sensor selection, evaluates data credibility using generalized and narrow Jaccard coefficients, and employs an improved AdaBoost classifier for intelligent scene recognition. By combining filtering credibility and data credibility, the proposed MCUKF enables adaptive fusion of multi-sensor data. Experimental results from simulations and real flight tests demonstrate its effectiveness. In the adaptive sensor fusion experiment, the proposed method achieves an RMSE of 0.1333 m, compared with 0.6403 m for GPS, 0.7697 m for the camera, and 0.4873 m for direct GPS-camera fusion. In comparison with representative fusion algorithms, MCUKF attains an RMSE of 0.1806 m, outperforming MMSE (0.5140 m), EUKF (1.3326 m), and STETFF (0.9568 m). These results demonstrate the superiority of the proposed method in adaptive sensor selection and robust fusion during scene transformation.