The importance of measurement uncertainty modelling in the feature-based RGB-D SLAM

This paper presents an analysis of the role of measurements uncertainty in the feature-based RGB-D SLAM formulated as graph optimization problem. The considered SLAM solution uses global graph optimization to find the trajectory of the RGB-D camera and a set of 3D point features constituting the map. In order to focus on the optimization back-end details and isolate the results from the data association errors caused by the image processing front-end in a real SLAM system we introduce a simulation environment, which allows to clearly show the influence of the uncertainty model on the accuracy of the obtained trajectories. We demonstrate a substantial improvement in the trajectory accuracy due to using in the graph optimization process an uncertainty model based on the physical properties of the RGB-D sensor. Moreover, we investigate the influence of the RGB-D camera motion strategy on the accuracy of the SLAM solution, pointing out the relation between this strategy and the measurement uncertainty model.