Analysis on Pose Estimation Uncertainty and Observation Parametrization for RGB-D Cameras

For the pose estimation based on RGB-D cameras, the pose estimation uncertainties of the ICP (iterative closest point) and PnP (perspective-n-point) algorithms are analyzed according to the maximum likelihood estimation and covariance propagation law, and the differences between the two algorithms in practical applications are explored from the perspective of data association. In addition, the influence of different parametrization forms of feature observation on the localization accuracy of visual odometer is compared, and a parametrization method is proposed. Based on the difference and connection between ICP and PnP algorithms, a switching strategy for the two PnP algorithms in the RGB-D SLAM (simultaneous localization and mapping) system is proposed, and different weights are assigned to each error term in the optimization problem according to the uncertainty of the observation. Experiments on two public datasets show that compared with the mainstream RGB-D SLAM algorithms, the proposed algorithm has higher localization accuracy and robustness in a variety of scenes with low texture, fast camera movement or dynamic objects. Meanwhile, the time efficiency of the proposed algorithm is about 10% higher than ORB-SLAM2 algorithm.