A novel adaptive keyframe selection method with multi-source joint constraints for visual SLAM

The effectiveness of keyframe selection is crucial for visual simultaneous localization and mapping (SLAM) algorithms. Stability and adaptability of the selection are challenging issues in complex environments. This paper introduces a novel Adaptive Keyframe Selection method with Multi-Source Joint Constraints (MSJCA-KS), which addresses the challenges by exploiting camera geometry, real-time inertial measurement unit (IMU) data analysis, and effective point distribution constraints. Particularly, we used adaptive threshold conditions based on camera geometry for the keyframe selection under varying scenes. A real-time state detection mechanism is used to refine these thresholds by using IMU data during rapid motion. Furthermore, a constraint mechanism is used to ensure a balanced distribution of keyframes. Our keyframe selection method is integrated into the ORB-SLAM3 framework to perform SLAM tasks. The effectiveness of our method is validated on the EuRoC dataset and the TUM dataset in terms of absolute trajectory error (ATE), keyframe quality, and computational efficiency. When compared to the photogrammetric keyframe selection (PKS) algorithm, the proposed method achieves a 16% improvement in localization accuracy in the EuRoC dataset and 27% and 23% in the TUM dataset for monocular and stereo cameras, respectively.