A vision-inertial interaction-based autonomous UAV positioning algorithm

Abstract In wartime scenarios where global navigation satellite system (GNSS) may be compromised, the cumulative errors associated with inertial navigation systems (INSs) can lead to significant positional deviations during long-distance navigation of drones. To address this issue, this study proposes an innovative vision-inertial interactive autonomous positioning algorithm that integrates visual positioning with inertial measurement unit-based methods, aiming to achieve enhanced system robustness and optimal positioning accuracy. The algorithm first implements an adaptive tile map level selection mechanism based on prior positional information from INS and onboard barometric pressure data, thereby mitigating the impact of inconsistent image scales. Subsequently, the proposed feature point extraction algorithm (Super Point V) effectively filters out invalid and interfering feature points from aerial images and tile maps. Finally, a novel matching mechanism for map expansion based on direction weights is employed to facilitate image matching and positioning, ensuring real-time accuracy and precision. Experimental results across various terrain conditions demonstrate that this approach not only enables accurate real-time computation of latitude and longitude but also effectively corrects the accumulated positioning errors of INS, achieving an average positioning accuracy error of only 4.94 m, comparable to that of conventional civilian GNSS, while maintaining excellent robustness.