基于DSG-SLAM的无人机城市低空环境感知方法OA

To address the challenges of dynamic object detection and obstacle avoidance in complex urban low-altitude environments,a dynamic semantic-geometric SLAM(DSG-SLAM)framework for unmanned aerial vehicle(UAV)perception is proposed.The method integrates semantic segmentation and geometric features to improve dynamic scene understanding.First,color and depth images are input into the system,where YOLOv11 and BoT-SORT algorithms are employed for object detection and tracking,enabling the discrimination between static environments and dynamic targets.Then,for each detected dynamic target region,point cloud generation,Euclidean clustering,and axis-aligned bounding box construction are performed to compute the size,position,and velocity of the objects.Finally,the static features are extracted and matched to estimate UAV pose and build a static octree-based map.The system outputs multimodal perception information,including UAV pose,static environmental structure,and the position and velocity of dynamic objects.Experiments on a custom-built urban dataset show that our method achieves a pose estimation error below 1％,a dynamic object localization error below 5％,and real-time performance at 20 FPS.These results confirm the effectiveness of our approach for UAV perception in complex urban environments.