A Visual-Inertial Pressure Fusion-Based Underwater Simultaneous Localization and Mapping System

Zhufei Lu, Xing Xu, Yihao Luo, Lianghui Ding, Chao Zhou, Jiarong Wang2024 · DOI

TL;DR

This paper introduces ORB-SLAM3-VIP, a novel system that fuses visual, inertial, and pressure sensor data for more accurate underwater Simultaneous Localization and Mapping (SLAM). By incorporating depth constraints from a pressure sensor, the system significantly improves scale estimation and trajectory accuracy compared to existing visual-inertial SLAM methods, crucial for autonomous underwater navigation.

What Problem Does This Paper Address?

Navigating underwater autonomously is a significant challenge. While visual SLAM systems excel on land, the underwater environment presents unique difficulties: suspended particles cause blurring, and light and color attenuate, leading to noisy and unreliable visual features. This noise can cause visual SLAM systems to lose track or produce inaccurate maps. Existing visual-inertial SLAM systems, while better, still struggle with accurate scale estimation and can suffer from cumulative drift. Detecting objects like naval mines requires high-precision navigation, which current single-sensor or even visual-inertial systems cannot reliably provide in complex underwater conditions. The paper aims to bridge this gap by developing a more robust and accurate SLAM system for underwater environments.

Methods

The core of this work is the extension of the ORB-SLAM3 framework, specifically the visual-inertial version (ORB-SLAM3-VI), by integrating a pressure sensor. The proposed system, ORB-SLAM3-VIP, achieves this through a "tight coupling" approach, meaning the data from the visual, inertial, and pressure sensors are jointly optimized within a factor graph framework.

Key methodological contributions include:

  • Multi-Sensor Fusion: Integrating depth measurements from a pressure sensor alongside visual and IMU data.
  • Depth Constraints: Introducing depth constraints into the SLAM optimization process. These constraints are used to:
  • Initialization: Obtain better initial estimates for scale, gravity direction, and IMU biases.
  • Scale Fine-Tuning: Correct the scale of the reconstructed map to match the real world.
  • Tracking and Mapping: Improve the accuracy of pose estimation and map generation by constraining the sensor's position along the z-axis (depth).
  • Factor Graph Optimization: Utilizing a factor graph to represent the probabilistic relationships between sensor measurements and state variables (pose, velocity, IMU biases). This allows for efficient optimization of all sensor data simultaneously.
  • Data Pretreatment: Implementing soft synchronization for multi-sensor data, mapping IMU measurements to camera frames and handling varying pressure sensor frequencies through averaging or interpolation.

Datasets

The paper evaluates its proposed system on the AQUALOC dataset. This dataset is specifically designed for underwater SLAM and includes sequences with visual, inertial, and depth (pressure) sensor data. The authors use seven sequences from this dataset for their experiments.

Key Insights and Ideas

  • Pressure Sensor as a Game-Changer for Scale: The most significant insight is how a simple pressure sensor, providing depth information, can fundamentally solve the scale ambiguity problem inherent in monocular visual SLAM. Unlike IMUs, which provide relative motion but are prone to drift and don't directly constrain absolute position, pressure sensors offer a direct, albeit noisy, measure of depth.
  • Tight Coupling is Crucial: Simply fusing data isn't enough; the paper emphasizes the importance of "tightly coupling" the sensors within the optimization framework. This allows the system to leverage the complementary strengths of each sensor to correct the weaknesses of others.
  • Depth Constraints Improve Z-Axis Accuracy: By explicitly incorporating depth constraints, the system gains a strong anchor for the vertical (z-axis) component of its pose estimation, which is often the most challenging in underwater environments due to the lack of distinct visual features at different depths.
  • Faster and More Accurate Scale Recovery: The ORB-SLAM3-VIP system converges to the correct scale much faster and more accurately than its visual-inertial counterpart, as demonstrated by the experimental results.

Key Results

  • Scale Error Reduction: The ORB-SLAM3-VIP system reduces scale errors by up to 41.2% compared to ORB-SLAM3-VI across all tested sequences.
  • Trajectory Error Reduction: The system also reduces trajectory errors by up to 41.2%.
  • Overall Accuracy Improvement: The square root of the trajectory error is reduced by up to 41.6%.
  • Improved Depth Estimation: Visualizations show that the z-axis estimation (depth) of ORB-SLAM3-VIP is significantly closer to the true value compared to ORB-SLAM3-VI.
  • Robustness: The system demonstrates improved robustness, particularly in scenarios where visual tracking might be lost for short periods.

Limitations and Future Work

  • Limitations:
    • The paper acknowledges that while the pressure sensor helps with scale, the system still relies on visual features for tracking. If visual features are severely degraded (e.g., in very murky water or during prolonged periods of no visual input), performance can still be affected.
  • The performance is evaluated on the AQUALOC dataset, which is specific to certain underwater conditions. Generalizability to all underwater environments might require further testing.
  • Future Work:
    • The authors suggest that this framework can be extended to improve the operational efficiency of naval mine countermeasures, implying potential for object detection and mapping in MCM operations.
  • Further exploration of more advanced sensor fusion techniques or alternative depth sensors could be a direction.
  • Investigating methods to handle extreme visual degradation more gracefully, perhaps by relying more heavily on IMU and pressure data during such periods, could be beneficial.

Summary updated

BibTeX
@misc{lu2024visual,
  author = {Zhufei Lu and Xing Xu and Yihao Luo and Lianghui Ding and Chao Zhou and Jiarong Wang},
  title = {{A Visual-Inertial Pressure Fusion-Based Underwater Simultaneous Localization and Mapping System}},
  year = {2024},
  doi = {10.3390/s24103207},
}

Loading related content...

Figure