Front-End Optimization Algorithm for Underwater Visual Simultaneous Localization and Mapping

2026-04-01

PRE

Ming, Liu

Hao, Qian *

DOI：10.3788/LOP252012

Original

Original request for help

Save

Abstract

En 中文

Objective Visual simultaneous localization and mapping (VSLAM) serves as a fundamental technology for autonomous robot navigation, particularly in GPS-denied environments such as underwater scenarios. However, the practical application of VSLAM in underwater environments faces significant challenges due to complex optical conditions, including low contrast, color distortion, and suspended particle interference. These factors severely degrade image quality, leading to insufficient robustness in feature extraction and high mismatch rates in existing SLAM systems such as UFEN-SLAM. The limitations of current approaches are particularly evident in their inability to adapt to non-uniform illumination and turbid conditions, resulting in feature detection failures in low-texture regions and redundant features in high-noise areas. This research aims to develop an optimized front-end algorithm specifically designed for underwater VSLAM, focusing on enhancing feature extraction and matching mechanisms to improve localization accuracy and system robustness in challenging underwater conditions. The proposed algorithm seeks to overcome the domain shift problems associated with deep learning methods while addressing the limitations of traditional indirect methods in handling severe image degradation. Methods The proposed front-end optimization algorithm comprises two main innovative stages that work synergistically to enhance performance. In the first stage, adaptive feature extraction is achieved through a novel global-local synergistic dynamic threshold mechanism. This mechanism intelligently combines image entropy, gradient magnitude, and local maximum between-class variance to construct an adaptive threshold that responds effectively to non-uniform degradation in underwater images. The process is further refined through quadtree-based spatial density control, which ensures uniform distribution of feature points across the image, and local contrast illumination stability filtering, which enhances robustness against lighting variations. The second stage focuses on geometric consistency matching and optimization, where we implement a sophisticated two-stage verification strategy. This includes rigorous bidirectional consistency checks followed by RANSAC epipolar constraints, which collectively work to eliminate erroneous matches effectively. The algorithm is implemented within the ORB-SLAM3 framework and comprehensively evaluated using multiple public datasets (AQUALOC, EASI, EuRoC) along with a self-collected dataset from Fuxian Lake in Yunnan, China, covering diverse underwater conditions including varying illumination, texture characteristics, and turbidity levels. Results and Discussions Extensive experimental results consistently demonstrate the superior performance of the proposed algorithm across multiple evaluation metrics. In feature extraction tests conducted under a strict limit of 500 keypoints, our algorithm extracted substantially more features while maintaining excellent spatial distribution quality. Particularly noteworthy is the performance on the challenging EASI dataset, where the proposed method extracted 471 features compared to merely 6 by ORB-SLAM3 and 200 by UFEN-SLAM. Quantitative analysis reveals significant improvements in spatial distribution metrics, with average spatial entropy reaching 0.849, representing improvements of approximately 42.2 % and 4.3 % over ORB-SLAM3 and UFEN-SLAM, respectively. Similarly, the grid coverage ratio showed improvements of approximately 39.1 % and 11.9 % over the same baseline algorithms. Visual comparisons of feature extraction results across different datasets clearly illustrate the enhanced distribution and quantity of detected features. In feature matching evaluation, the proposed algorithm demonstrated remarkable accuracy and robustness. The algorithm consistently produced more matched pairs while achieving higher mean matching accuracy across all test datasets, reaching an exceptional 0.997 on the in-house dataset. The matching process, though slightly more computationally expensive than ORB-SLAM3, provided significantly better accuracy than all comparison algorithms. The effectiveness of our matching optimization strategy is visually evident in Fig. 4 and Fig. 5, which show denser and more accurate matching results after optimization. For localization performance, comprehensive tests on the EuRoC dataset sequences confirmed the superior accuracy of our approach. The algorithm achieved the lowest absolute trajectory error on most sequences, with a particularly impressive RMSE of 0.009 m on V102, representing an approximate 18% improvement over ORB-SLAM3. The trajectory comparison shown in Fig. 6 visually demonstrates the closer alignment of our algorithm's estimated trajectory with the ground truth. The system also exhibited enhanced computational efficiency, with running times significantly reduced compared to both UFEN-SLAM and ORB-SLAM3 across all test sequences. Detailed trajectory and attitude error analyses provided further evidence of the improved stability and accuracy of the proposed method. Additional mapping experiments conducted in authentic underwater scenarios validated the system's capability for real-time localization and sparse point cloud construction, as clearly demonstrated in Fig. 11, Fig. 12, and Fig. 13. The consistent performance improvement across all evaluation dimensions confirms that our front-end optimization strategy effectively addresses the specific challenges of underwater visual SLAM. Conclusions This paper presents a comprehensively optimized front-end algorithm for underwater visual SLAM that effectively addresses the critical issues of low robustness in feature extraction and high mismatch rates encountered by UFEN-SLAM in complex underwater environments. The innovative integration of a global-local dynamic threshold mechanism for feature detection, combined with sophisticated spatial density and illumination stability filtering, ensures the generation of a balanced and highly reliable feature set that adapts well to challenging underwater conditions. The implementation of a two-stage geometric verification strategy, incorporating bidirectional consistency checks and RANSAC epipolar constraints, provides robust elimination of mismatches while maintaining computational efficiency. Extensive experimental validation on multiple datasets confirms that the proposed algorithm delivers substantial improvements in feature extraction quality, matching accuracy, real-time performance, and overall system robustness. The algorithm provides an effective and reliable solution for real-time localization and mapping in complex underwater environments, demonstrating practical application potential for underwater autonomous navigation systems. Future research will explore the applicability of this approach to other challenging visual conditions and SLAM frameworks utilizing different descriptor types, with cross-framework experimental validation planned as subsequent work.

Keywords:

visual simultaneous localization and mapping

feature extraction

feature matching

underwater environment

mismatch removal

Front-End Optimization Algorithm for Underwater Visual Simultaneous Localization and Mapping

Abstract

Journal

Organization