期刊名称：Journal of Automation, Mobile Robotics & Intelligent Systems (JAMRIS)
印刷版ISSN：1897-8649
电子版ISSN：2080-2145
出版年度：2019
卷号：13
期号：2
页码：5-14
DOI：10.14313/JAMRIS/2-2019/13
出版社：Industrial Research Inst. for Automation and Measurements, Warsaw
摘要：Visual odometry esmates the transformaons betweenconsecuve frames of a video stream in order to recoverthe camera’s trajectory. As this approach does notrequire to build a map of the observed environment, itis fast and simple to implement. In the last decade RGBDcameras proliferated in robocs, being also the sensorsof choice for many praccal visual odometry systems. AlthoughRGB-D cameras provide readily available depthimages, that greatly simplify the frame-to-frame transformaons computaon, the number of numerical parametersthat have to be set properly in a visual odometrysystem to obtain an accurate trajectory esmate remainshigh. Whereas seng them by hand is certainly possible,it is a tedious try-and-error task. Therefore, in this arclewe make an assessment of two populaon-based approachesto parameter opmizaon, that are for long meapplied in various areas of robocs, as means to find bestparameters of a simple RGB-D visual odometry system.The opmizaon algorithms invesgated here are par-cle swarm opmizaon and an evoluonary algorithmvariant. We focus on the opmizaon methods themselves,rather than on the visual odometry algorithm, seekingan efficient procedure to find parameters that minimizethe esmated trajectory errors. From the experimentalresults we draw conclusions as to both the efficiencyof the opmizaon methods, and the role of par-cular parameters in the visual odometry system.
