Združevanje vizualnega prepoznavanja lokacij in metode SLAM s histogramskim filtrom za robustno vizualno lokalizacijo

This thesis introduces a new approach to drone localization in GNSS-denied environments, improving an existing Visual Place Recognition (VPR)-based localization system by incorporating data about the relative motion of the drone. The relative motion data acquired from visual SLAM or odometry present the relative displacement of the drone between frames without a global reference, while the VPR systems provide globally consistent but often ambiguous and noisy location estimates. We propose a method that fuses the global location likelihoods with motion information within the Bayesian histogram filtering framework. The method is evaluated on the UAV-VisLoc dataset, which includes diverse environments and enables consistent comparison with the baseline VPR method. The results show that the proposed fusion method provides stable performance and improves localization accuracy in nearly all evaluated scenarios. This highlights the effectiveness of integrating motion data to enhance the VPR-based localization system’s accuracy, encouraging further research on similar methods.