A new keyframe decision mechanism with translation constraint for visual slam

Since the existing keyframe detection methods always ignore the translation information among frames, this paper proposes a new keyframe decision mechanism by combining this translation infomation. First, we obtain the translation information among the current two consecutive frames and design a translation constraint. Next, we refine five constraints: translation information, time interval, proportion of close and far point, idleness of local mapping and tracked points, in the meantime, skillfully combine them to form a new keyframe decision mechanism. Compared with the keyframe detection methods in ORBSLAM2, PTAM, LSDSLAM and RGBDSLAM, the visual SLAM with our keyframe decision mechanism obtains a smaller Root Mean Square Error (RMSE) and better localization, which is confirmed on four TUM data sequences, ten KITTI data sequences and the real-time laboratory environment.