On automatic parameter optimization in an RGB-D visual localization system

Abstract

Reliable visual localization is of high importance in many practical applications. Though RGB-D Visual Odometry is commonly used in applications and as a basis for more complicated Simultaneous Localization and Mapping systems, the diversity of details in the known solutions makes it hard to tell how particular design choices and parameter values influence the performance. Therefore, in this paper we investigate whenever it is possible to automatize the selection of parameters for a simple, feature-based RGB-D Visual Odometry system. We assume a fixed structure of the Visual Odometry pipeline and employ a population-based optimization algorithm to find the best parameters. The experiments are performed using publicly available datasets to ensure that the results are verifiable.