Improved Bag-of-Words Image Retrieval with Geometric Constraints for Ground Texture Localization

TL;DR

This work tackles robust BoW-based localization for ground texture using a downward-facing camera, addressing the need for accurate global localization and reliable loop closures in SLAM. It introduces four key enhancements—size-based feature binning, orientation-consistent verification, an AKM vocabulary, and soft descriptor assignment—and provides two optimized variants: a high-accuracy version for real-time global localization and a high-speed version for loop-closure detection. Ablation and timing studies demonstrate substantial gains over baselines like DBoW, with the high-accuracy variant achieving superior global localization and near-perfect loop-closure recall, while the high-speed variant offers scalable performance. The approach enables immediate integration into existing BoW-based ground texture pipelines, improving robustness in dynamic environments, and the authors release code for community use.

Abstract

Ground texture localization using a downward-facing camera offers a low-cost, high-precision localization solution that is robust to dynamic environments and requires no environmental modification. We present a significantly improved bag-of-words (BoW) image retrieval system for ground texture localization, achieving substantially higher accuracy for global localization and higher precision and recall for loop closure detection in SLAM. Our approach leverages an approximate $k$-means (AKM) vocabulary with soft assignment, and exploits the consistent orientation and constant scale constraints inherent to ground texture localization. Identifying the different needs of global localization vs. loop closure detection for SLAM, we present both high-accuracy and high-speed versions of our algorithm. We test the effect of each of our proposed improvements through an ablation study and demonstrate our method's effectiveness for both global localization and loop closure detection. With numerous ground texture localization systems already using BoW, our method can readily replace other generic BoW systems in their pipeline and immediately improve their results.

Figures (5)

• Figure 1: Our algorithm converts images of the ground texture into a bag-of-words representation to localize a mobile robot.
• Figure 2: Our proposed BoW method. Image features are extracted from the ground image and each descriptor is assigned to a visual word via an AKM vocabulary. Additionally, the size of each feature is binned into one of $S$ pre-trained bins. Then, when inserting entries into the inverse index, each element is inserted into the row that corresponds to that feature's vocabulary word and size. Each entry consists of an image ID, the word's weight in the BoW vector, and a list of features that correspond to the word in that image. We store each feature's ID, orientation, and contributing weight to later use for orientation verification.
• Figure 3: The database insertion and querying timings as the size of the database increases.
• Figure 4: Global localization performance of different algorithms on the HD Ground asphalt dataset over time hdground. The bold vertical line indicates the database recording date, the dashed lines indicate test set recording dates.
• Figure 5: Top Left: The Recall@$N$ for DBoW and our algorithms on the asphalt dataset. Other: BoW similarity scores for potential loop closures for DBoW and our algorithms.