Sonar Context

[ICRA 2023] Robust Imaging Sonar-based Place Recognition and Localization in Underwater Environments
Hogyun Kim, Gilhwan Kang, Seokhwan Jeong, Seungjun Ma, Younggun Cho · SPARO Lab, Inha University

Underwater there is no GPS and no light; a forward-looking imaging sonar sees 50 m ahead through murky water, but its images are noisy, low-resolution and change with the viewpoint. Sonar Context turns one sonar image into a small, robust place descriptor, so a robot can tell that it has been here before.

• 🌊 Sonar Context: the polar sonar image pooled into patches (range rings × bearing sectors), plus a polar key (one value per range ring) that does not change with the heading.
• 🔍 Coarse to fine: the polar key finds candidates in a k-d tree; the Sonar Context compares them column by column over every bearing shift (a rotation) and row by row (a motion along the beams).
• 🔁 SLAM: DISO sonar odometry + Sonar Context loops, verified by dense phase correlation and optimized with iSAM2, on ROS 1 and ROS 2.
• 👥 Multi-session and multi-robot: sessions recorded apart are aligned through their shared places; two robots merge their maps with Commerge from a few keyframes.
• 📦 Library first: pip install sonar-context (NumPy), a C++17 library with the same results, and every application built on them.

📰 News

• [Nov 2026] v2.0.0: Python and C++ libraries, SLAM with DISO (ROS 1 / ROS 2), multi-session SLAM, multi-robot SLAM with Commerge, and fixes to the matching (below). The ICRA 2023 code is kept at v1.0.0.
• [Jan 2023] Sonar Context is accepted to ICRA 2023.

🛠️ Installation

pip install sonar-context                 # Python 3.9+, NumPy / SciPy / Pillow
git clone --recursive https://github.com/sparolab/sonar_context.git   # C++, SLAM (DISO as a submodule)

Python

import sonar_context as sc

# a polar sonar image: rows = range bins, columns = beams (layout="beam_range" if stored the other way)
desc = sc.sonar_context(image, patch=(4, 4))
dist, shift = sc.column_distance(desc_a, desc_b)        # 1 - similarity, best bearing shift (a rotation)

m = sc.SonarContextManager(col_threshold=0.2, exclude_recent=30)
for image in sequence:
    loop = m.update(image)                               # Loop(query, match, distance, column_shift, ...) or None

sonar-context describe polar_image.zip --layout beam_range --out seq.npz      # descriptors of a sequence
sonar-context loops seq.npz --out loops.csv --poses poses.txt               # online loop detection (+ precision)
sonar-context affinity a.npz b.npz --out affinity.npy                       # two sessions, for Commerge

C++

sudo apt install libeigen3-dev
cmake -S cpp -B cpp/build && cmake --build cpp/build -j && ctest --test-dir cpp/build

#include "sonar_context/sonar_context.hpp"
auto d = sonar_context::make_sonar_context(image);                 // CMake: sonar_context::sonar_context
auto m = sonar_context::column_distance(d_a, d_b);                 // m.distance, m.shift

Same distances and shifts as Python (checked in python/tests/test_cpp.py).

🗺️ SLAM

applications/slam: any odometry (default DISO, direct imaging sonar odometry, ICRA 2024) gives keyframes every 1 m; Sonar Context proposes revisits; each is verified by dense phase correlation of the two sonar frames (rotation searched, normalized peak ≥ 0.10) and closed with iSAM2. The core is ROS-free; the nodes are thin wrappers.

docker build -f docker/Dockerfile.slam -t sonar-context-slam:noetic .            # ROS 1: DISO + SLAM
docker run --rm -v <Aracati>:/data:ro -v $PWD/out:/out sonar-context-slam:noetic \
    roslaunch sonar_context_slam aracati2017.launch bag:=/data/ARACATI_2017_8bits_full.bag save_dir:=/out/slam \
    record:=/out/aracati_diso.bag                                                # + a bag for ROS 2
docker build -f docker/Dockerfile.slam-ros2 -t sonar-context-slam:humble .       # ROS 2: SLAM on recorded poses
docker run --rm -v $PWD/out:/out sonar-context-slam:humble \
    ros2 launch sonar_context_slam aracati2017.launch.py bag:=/out/aracati_diso_ros2 save_dir:=/out/slam_ros2

On Aracati 2017 (44 min, 620 m in a marina, BlueView P900-130), live with DISO, against the DGPS ground truth (evo_ape -a; scale-corrected -as in parentheses):

DISO is not deterministic from run to run (4.5 to 5.1 m here); the loops bring every run under 2 m. DISO's poses are in a z-down frame (flip_frame: true turns them, and the sonar beams, to z-up). Everything else: applications/slam.

👥 Multi-session SLAM

applications/multi-session-slam: LT-SLAM style (LT-mapper), as in RadLoc and ReFeree: two sessions saved by the SLAM keep their own frames and get one anchor node each; inter-session Sonar Context loops, verified by dense correlation, connect them through the anchors (Cauchy kernel).

sonar_context_align_sessions session_a/ session_b/ merged/ --threshold 0.25

Aracati 2017 split in two sessions: 47 inter-session loops (37 agree on one transform); the two sessions together come to 2.52 m APE (17.88 m with session B left in its own frame; 1.96 m and 2.33 m for each session alone).

🤖 Multi-robot SLAM with Commerge

applications/multi-robot-slam: two robots merge their maps by exchanging only a few keyframes, chosen by Commerge (IJRR 2026) from the Sonar Context affinity between them: on Aracati split into two robots, 4.5 % of the keyframes (2.6 KB of poses) merge the maps within 1.3 m.

🔧 What changed in v2.0.0

The matching of the ICRA 2023 code had bugs, fixed in both libraries (the old code: v1.0.0):

• the best bearing shift was read one entry off (argmax + 1), and could index past the array;
• shift 0 (the same heading) was never tested;
• the row search compared unshifted rows, so it had no effect;
• a shift with no non-empty overlap divided by zero.

New: skip_range drops the bright near field (on Aracati, 5 m), the polar image layout is explicit (range_beam / beam_range: Aracati's polar images are stored beams × range), and the descriptors of two sessions compare in one call (column_similarity_matrix, FFT-based).

🗃️ Datasets

• Aracati 2017: the bag (/son, /pose_gt) and the polar images; DISO's Aracati2017_DISO adds the dead-reckoning prior (thirdparty/).
• HoloOcean simulation: Google Drive.
• KRISO (Korea Research Institute of Ships and Ocean Engineering): not public.

Citation

@inproceedings{kim2023robust,
  title={Robust Imaging Sonar-based Place Recognition and Localization in Underwater Environments},
  author={Kim, Hogyun and Kang, Gilhwan and Jeong, Seokhwan and Ma, Seungjun and Cho, Younggun},
  booktitle={2023 IEEE International Conference on Robotics and Automation (ICRA)},
  pages={1083--1089},
  year={2023},
  organization={IEEE}
}

Acknowledgements

Sonar Context builds on Scan Context (Kim and Kim, IROS 2018); the multi-session SLAM follows LT-mapper (Kim et al., ICRA 2022). The SLAM uses DISO (Xu et al., ICRA 2024, GPLv3, a submodule) for odometry and GTSAM. Data: Aracati 2017 (dos Santos et al., RA-L 2022), HoloOcean (Potokar et al., ICRA 2022); multi-session underwater SLAM: Jang et al., ICRA 2021. Related: SOLiD, Commerge, ReFeree, RadLoc.

License

BSD 3-Clause for academic use; for any commercial purpose, please contact the authors (hg.kim@inha.edu). DISO (thirdparty/DISO) is GPLv3 and is not part of this license.