Dynamic Strategy of Keyframe Selection with PD Controller for VSLAM Systems

Abstract

Keyframe (KF) selection in a KF-based Visual Simultaneous Localization and Mapping (VSLAM) system is critical. In previous studies, static thresholds have been used for KF selection decision-making; however, suboptimal performance can result from the use of such thresholds. To obtain a better KF setting than that obtained with the existing methods, we introduce a dynamic KF selection strategy. By considering both the view change between camera observation and KFs in the built map and the rate of this change, we propose to dynamically adjust the threshold for KF selection. A Proportion and Derivative (PD) controller is designed with the feedback of estimated view change, where the PD controller output is used for KF selection. According to the experimental results, compared with the existing studies, our method can improve the precision of visual tracking by 17.5% and 16.7% based on two popular VSLAM systems.