Keyframe Selection for Visual Localization and Mapping Tasks: A Systematic Literature Review

Abstract

Visual localization and mapping algorithms attempt to estimate, from images, geometrical models that explain ego motion and the positions of objects in a real scene. The success of these tasks depends directly on the quality and availability of visual data, since the information is recovered from visual changes in images. Keyframe selection is a commonly used approach to reduce the amount of data to be processed as well as to prevent useless or wrong information to be considered during the optimization. This study aims to identify, analyze, and summarize the methods present in the literature for keyframe selection within the context of visual localization and mapping. We adopt a systematic literature review (SLR) as the basis of our work, built on top of a well-defined methodology. To the best of our knowledge, this is the first review related to this topic. The results show that there is a lack of studies present in the literature that directly address the keyframe selection problem in this application context and a deficiency in the testing and validation of the proposed methods. In addition to these findings, we also propose an updated categorization of the proposed methods on top of the well-discussed categories present in the literature. We believe that this SLR is a step toward developing a body of knowledge in keyframe selection within the context of visual localization and mapping tasks by encouraging the development of more theoretical and less heuristic methods and a systematic testing and validation process.

Visual Simultaneous Localization and Mapping Optimization Method Based on Object Detection in Dynamic Scene

“…In 2021, they even unveiled ORB-SLAM3 with a multi-map system that incorporates a feature-based inertial visual odometry that greatly enhances the system's localization accuracy [9]. However, there are still certain issues with ORB-SLAM3 in practical situations [10]. While moving objects in actual realworld environments can cause errors in the correlation of visual odometry data, existing algorithms typically assume the external environment to be static [11].…”