Robust Imaging Sonar-based Place Recognition and Localization in Underwater Environments

TL;DR

The paper tackles robust place recognition and loop closure for underwater SLAM using imaging SONAR without supervision. It introduces SONAR Context and Polar Key descriptors, plus adaptive shifting and padding to handle rotation and translation, integrating ICP-based loop closures in a pose-graph SLAM framework. The approach is validated across HOLOOCEAN, KRISO water tank, and ARACATI datasets, showing superior precision/recall and reduced drift compared to baselines, including feature-based and prior global-descriptor methods. This work offers a practical, real-time-friendly solution for robust underwater localization and mapping, with potential extensions to other SONAR modalities and semantic-informed representations.

Abstract

Place recognition using SOund Navigation and Ranging (SONAR) images is an important task for simultaneous localization and mapping(SLAM) in underwater environments. This paper proposes a robust and efficient imaging SONAR based place recognition, SONAR context, and loop closure method. Unlike previous methods, our approach encodes geometric information based on the characteristics of raw SONAR measurements without prior knowledge or training. We also design a hierarchical searching procedure for fast retrieval of candidate SONAR frames and apply adaptive shifting and padding to achieve robust matching on rotation and translation changes. In addition, we can derive the initial pose through adaptive shifting and apply it to the iterative closest point (ICP) based loop closure factor. We evaluate the performance of SONAR context in the various underwater sequences such as simulated open water, real water tank, and real underwater environments. The proposed approach shows the robustness and improvements of place recognition on various datasets and evaluation metrics. Supplementary materials are available at https://github.com/sparolab/sonar_context.git.

Figures (8)

• Figure 1: Our place recognition method using SONAR context, polar key and adaptive shifting. The figure on the left shows qualitative evaluation including trajectory with current frame and loop candidates in the ARACATI 2017 dataset dos2022cross. Candidates are selected through polar key, and adaptive shifting is applied between SONAR contexts to match. More details of our method are shown in Fig. \ref{['fig:our_slam_figure']}.
• Figure 2: Two types of typical SONAR images. A polar image of area (a) is mapped to a corresponding area of the encoded polar image (b). The encoded image includes range($r$) and azimuth($\theta$).
• Figure 3: Our proposed loop closure detection pipeline. Place description and point cloud processing are conducted in parallel. The place description part defines the SONAR context and polar key. Place recognition finds the candidate via polar key, applies adaptive shifting, and compares cosine similarity between the query and candidate. Finally, loop closing is achieved using ICP.
• Figure 4: Time-elevation trajectory with correct(green) and incorrect(red) matching for various methods and their precision-recall curves including self-ablated methods. We show these qualitative results when each method has the highest possible level of precision as much as possible. In detail, AKAZE+p is the method achieved by global image search with a polar key and the distance based on the AKAZE inlier, whereas AKAZE is the method achieved by one-by-one feature matching with no polar key involved.
• Figure 5: Histogram of detected loop pairs by rotation and translation difference. The figure is plotted at the recall is 0.4 for each method.