Research on UAV Location Algorithm Based on Improved ORBSLAM3 in Outdoor Environment

Visual SLAM algorithms are widely used in unmanned intelligent systems. However, most SLAM algorithms are based on the assumption of static environments, where there are various dynamic objects in the real world and dynamic changes in the overall environment; meanwhile, VSLAM systems drift in large environments, which reduce the accuracy of attitude estimation and localization. To address above problems, this paper proposes an improved ORBSLAM3 algorithm, which aims to provide accurate attitude estimation and global localization for autonomous UAVs in outdoor scenes. Building upon ORBSLAM3, firstly, a method combining the YOLOv5s detection algorithm with geometric verification is proposed to eliminate dynamic features. Secondly, a drift correction strategy is developed to address the attitude drift in larger environments. This strategy involves loosely coupling GPS absolute measurements with the ORBSLAM3 system, enabling the acquisition of global observations. Thirdly, we test the two improvements separately on the dynamic TUM dataset and the UAV EuRoC dataset, and both improvements show significant enhancements compared to ORBSLAM3. Finally, the overall improved ORBSLAM3 algorithm is tested on the KITTI dataset in outdoor scenarios, resulting in a reduction of over 70% in absolute trajectory error and achieving high localization accuracy.