At a Glance


Browse Publications by Ryan Eustice and the rest of the PeRL Team.

Browse by year

2018, 2017, 2016, 2015, 2014, 2013, 2012, 2011, 2010, 2009, 2008, 2007, 2006, 2005, 2004, 2003, 2002, 2000


Place recognition and localization for multi-modal underwater navigation with vision and acoustic sensors


Jie Li, Place recognition and localization for multi-modal underwater navigation with vision and acoustic sensors. PhD thesis, Department of Electrical Engineering and Computer Science, University of Michigan, August 2017.


Place recognition and localization are important topics in both robotic navigation and computer vision. They are a key prerequisite for simultaneous localization and mapping (SLAM) systems, and also important for long-term robot operation when registering maps generated at different times. The place recognition and relocalization problem is more challenging in the underwater environment because of four main factors: 1) changes in illumination; 2) long-term changes in the physical appearance of features in the aqueous environment attributable to biofouling and the natural growth, death, and movement of living organisms; 3) low density of reliable visual features; and 4) low visibility in a turbid environment. There is no one perceptual modality for underwater vehicles that can single-handedly address all the challenges of underwater place recognition and localization.

This thesis proposes novel research in place recognition methods for underwater robotic navigation using both acoustic and optical imaging modalities. We develop robust place recognition algorithms using both optical cameras and a Forward-looking Sonar (FLS) for an active visual SLAM system that addresses the challenges mentioned above.

We first design an optical image matching algorithm using high-level features to evaluate image similarity against dramatic appearance changes and low image feature density. A localization algorithm is then built upon this method combining both image similarity and measurements from other navigation sensors, which enables a vehicle to localize itself to maps temporally separated over the span of years.

Next, we explore the potential of FLS in the place recognition task. The weak feature texture and high noise level in sonar images increase the difficulty in making correspondences among them. We learn descriptive image-level features using a convolutional neural network (CNN) with the data collected for our ship hull inspection mission. These features present outstanding performance in sonar image matching, which can be used for effective loop-closure proposal for SLAM as well as multi-session SLAM registration. Building upon this, we propose a pre-linearization approach to leverage this type of general high-dimensional abstracted feature in a real-time recursive Bayesian filtering framework, which results in the first real-time recursive localization framework using this modality.

Finally, we propose a novel pose-graph SLAM algorithm leveraging FLS as the perceptual sensors providing constraints for drift correction. In this algorithm, we address practical problems that arise when using an FLS for SLAM, including feature sparsity, low reliability in data association and geometry estimation. More specifically, we propose a novel approach to pruning out less-informative sonar frames that improves system efficiency and reliability. We also employ local bundle adjustment to optimize the geometric constraints between sonar frames and use the mechanism to avoid degenerate motion patterns.

All the proposed contributions are evaluated with real-data collected for ship hull inspection. The experimental results outperform existent benchmarks. The culmination of these contributions is a system capable of performing underwater SLAM with both optical and acoustic imagery gathered across years under challenging imaging conditions.

Bibtex entry

@PHDTHESIS { jli-phdthesis,
    AUTHOR = { Jie Li },
    TITLE = { Place recognition and localization for multi-modal underwater navigation with vision and acoustic sensors },
    SCHOOL = { Department of Electrical Engineering and Computer Science, University of Michigan },
    YEAR = { 2017 },
    MONTH = { August },
    ADDRESS = { Ann Arbor, MI, USA },