Adaptive Keyframe Selection Strategy of Visual SLAM in Complex Poses

Aiming at the accuracy problem of pose estimation under complex poses of indoor unmanned aerial vehicles (UAVs), this article proposes a dynamic threshold setting method based on vision sensor pose detection for adaptive selection of keyframes. Due to the uncertainty of camera pose estimation, this article uses statistical theory to analyze the pose difference between camera frames to characterize the camera motion characteristics, innovatively employs a multilayered fuzzy inference mechanism for joint judgment to derive dynamic thresholds, and applies this method to ORB-SLAM3 system to construct a new system called fuzzy-ORB-SLAM3 for the purpose of achieving adaptive compensation in keyframe creation. Finally, the new system is validated on TUM and EuRoc public datasets. The results compared with ORB-SLAM3 and other typical SLAM systems show that the proposed system has higher positioning accuracy and can better adapt to the pose estimation of UAVs in complex motion states.