Performance Analysis of Feature Detectors and Descriptors in Underwater and Polar Environments

Many scientific mapping surveys that deploy robotic platforms in underwater and polar environments perform Visual Simultaneous Localization and Mapping (VSLAM), Structure for Motion (SfM), and Image Mosaicking. These techniques heavily rely on robust and reliable feature-based vision front ends. The job of a vision front end is to provide correspondence information between different camera views which is then directly fed into a bundle adjustment step. Although the atomic steps involved in constructing a vision front end are well-known, many of the popular choices of the features and their parameters do not perform reliably in a variety of visually degraded underwater and polar environments that are characterized by low texture and contrast, and by unevenly lit and low-overlap imagery. In this paper, we develop novel metrics and quantitative analysis methods which can measure the impact of image pre-processing steps such as Contrast Limited Adaptive Histogram Equalization (CLAHE) on the improvement of vision front-end outputs. Our metrics and quantitative analysis can guide the selection between different feature detectors and descriptors to develop a reliable and robust vision front end that can operate in a wider range of underwater and polar environments. We showcase how CLAHE improves the saliency of features and feature track length on a visually degraded dataset underwater, resulting in a substantial increase in correspondence information for the SfM solution. Finally, we perform an end-to-end SfM analysis that shows reduced accumulated drift over the long term and improved accuracy.