Localizing pre-visited places during long-term simultaneous localization and mapping, i.e. loop closure detection (LCD), is a crucial technique to correct accumulated inconsistencies. As one of the most effective and efficient solutions, Bag-of-Words (BoW) builds a visual vocabulary to associate features and then detect loops. Most existing approaches that build vocabularies off-line determine scales of the vocabulary by trial-and-error, which often results in unreasonable feature association. Moreover, the accuracy of the algorithm usually declines due to perceptual aliasing, as the BoW-based method ignores the positions of visual features. To overcome these disadvantages, we propose a natural convergence criterion based on the comparison between the radii of nodes and the drifts of feature descriptors, which is then utilized to build the optimal vocabulary automatically. Furthermore, we present a novel topological graph verification method for validating candidate loops so that geometrical positions of the words can be involved with a negligible increase in complexity, which can significantly improve the accuracy of LCD. Experiments on various public datasets and comparisons against several state-of-the-art algorithms verify the performance of our proposed approach.

中文翻译：

---

用于准确环路闭合检测的自动词汇表和图形验证

在长期同步定位和映射期间定位预先访问过的地方，即回环检测 (LCD)，是纠正累积不一致的关键技术。作为最有效和最高效的解决方案之一，词袋 (BoW) 构建了一个视觉词汇表来关联特征，然后检测循环。大多数离线构建词汇表的现有方法通过反复试验确定词汇量的规模，这通常会导致不合理的特征关联。此外，算法的准确性通常会由于感知混叠而下降，因为基于 BoW 的方法忽略了视觉特征的位置。为了克服这些缺点，我们提出了一种基于节点半径和特征描述符漂移之间的比较的自然收敛标准，然后用于自动构建最佳词汇表。此外，我们提出了一种新的拓扑图验证方法来验证候选循环，从而可以涉及单词的几何位置，而复杂度的增加可以忽略不计，这可以显着提高 LCD 的准确性。对各种公共数据集的实验以及与几种最先进算法的比较验证了我们提出的方法的性能。