Improving Visual Feature Extraction in Glacial Environments

Steven D. Morad, Jeremy Nash, Shoya Higa, Russell Smith, Aaron Parness, Kobus Barnard

Abstract

Navigating glacial environments is a significant challenge for autonomous robots because snow and ice often appear featureless in the visible spectrum, causing visual odometry and SLAM algorithms to fail. This study demonstrates that near-infrared (NIR) imaging reveals hidden spatial structures related to ice grain size, significantly increasing the quantity and quality of trackable visual features. By utilizing the sensitivity of NIR albedo to ice grain surface area, the researchers show that reliable navigation is possible even in seemingly uniform glacial interiors.

Introduction

Glaciers are high-priority targets for scientific exploration, yet they remain incredibly dangerous for human researchers due to hidden crevasses and extreme weather. While autonomous robots are ideal for these environments, they suffer from "perception blindness." Standard computer vision algorithms rely on distinct features—corners or textures—to track movement. In glacial settings, the visible spectrum often presents a uniform white void, leading to massive "pose drift" where the robot loses track of its position. This research investigates whether switching to the near-infrared spectrum can reveal the textures necessary for robust navigation.

Experimental Methodology

The researchers hypothesized that the optical properties of ice change significantly in the near-infrared spectrum, specifically regarding Specific Surface Area (SSA)—the ratio of surface area to volume of ice grains. To test this, they developed a custom handheld camera rig consisting of:

  • Dual Cameras: Two identical PointGrey FLEA-3 monochrome cameras in a stereo configuration.
  • Optical Filters: One camera captured visible light, while the other utilized a filter wheel with NIR longpass filters (800nm to 1000nm).
  • Active Illumination: A 75W halogen-tungsten lamp provided consistent light across both spectrums.
  • Ground Truth: A VectorNav VN-200 IMU provided high-accuracy orientation data to validate visual estimates.

Images were processed using industry-standard feature extractors (SIFT, SURF, and FAST) after being normalized via Contrast Limited Adaptive Histogram Equalization (CLAHE).

Theoretical Basis and Key Insights

The core insight of this work is that NIR albedo (reflectivity) is far more sensitive to ice grain size than visible light. A patch of snow that appears uniform to the human eye actually contains a complex map of varying grain sizes visible in NIR.

Figure 5: NIR albedo sensitivity to ice grain radius across different wavelengths.

As shown in the figure above, albedo decreases as grain size increases, with the effect becoming more pronounced at higher wavelengths. The researchers identified 850nm as the "sweet spot," balancing the revelation of hidden features with sufficient light levels for low-noise imaging. Furthermore, because standard silicon-based sensors are naturally sensitive to NIR, this approach is highly accessible; researchers can simply remove standard IR-blocking filters and replace them with longpass filters.

Experimental Results

The study utilized a custom dataset from Igloo Cave at Mt. St. Helens, covering featureless firn walls, striated layers, and planar snow. The results were definitive:

  • Feature Quantity: In "Planar Snow" scenes, NIR imagery produced orders of magnitude more features than visible light.
  • Orientation Accuracy: The Valid Orientation Percentage (VOP)—the frequency with which visual features correctly estimated rotation—was significantly higher in NIR for difficult scenes.
  • Visual Contrast: NIR successfully revealed "melt-freeze crusts" (layers of refrozen water with large grains) that appeared as dark streaks in NIR but were invisible in visible light.
Figure 3: Side-by-side comparison of visible light (left) and NIR (right) imagery of a firn wall.

Conclusion and Future Work

While NIR imaging provides a breakthrough for glacial perception, limitations remain. The current data was collected in a cave with artificial lighting; performance in high-glare outdoor sunlight is still being evaluated. Additionally, NIR filters block significant light, requiring higher camera gain which can introduce noise. Future work will focus on integrating these NIR features into full SLAM pipelines to test long-term drift and exploring advanced Indium-Gallium-Arsenide (InGaAs) sensors for even deeper infrared sensitivity.

Summary updated

Abstract

Glacial science could benefit tremendously from autonomous robots, but previous glacial robots have had perception issues in these colorless and featureless environments, specifically with visual feature extraction. This translates to failures in visual odometry and visual navigation. Glaciologists use near-infrared imagery to reveal the underlying heterogeneous spatial structure of snow and ice, and we theorize that this hidden near-infrared structure could produce more and higher quality features than available in visible light. We took a custom camera rig to Igloo Cave at Mt. St. Helens to test our theory. The camera rig contains two identical machine vision cameras, one which was outfitted with multiple filters to see only near-infrared light. We extracted features from short video clips taken inside Igloo Cave at Mt. St. Helens, using three popular feature extractors (FAST, SIFT, and SURF). We quantified the number of features and their quality for visual navigation by comparing the resulting orientation estimates to ground truth. Our main contribution is the use of NIR longpass filters to improve the quantity and quality of visual features in icy terrain, irrespective of the feature extractor used.
BibTeX
@article{morad2019improving,
  author = {Steven D. Morad and Jeremy Nash and Shoya Higa and Russell Smith and Aaron Parness and Kobus Barnard},
  title = {{Improving Visual Feature Extraction in Glacial Environments}},
  journal = {IEEE Robotics Autom. Lett.},
  year = {2019},
  doi = {10.1109/lra.2019.2959490},
}
Figure