Bathymetric Feature Extraction and Matching on Semi-Synthetic Sonar Imagery

Feature extraction and matching is critical in land-mark association to aid autonomous underwater vehicle (AUV) navigation. Typical feature extraction methods are designed for optical imagery and are not optimized for underwater bathymetry. In this paper, a method of generating semi-synthetic bathymetry sonar images is presented, as a precursor to de-risk sea trials. These sonar image datasets are used to evaluate feature matching performance of feature detectors and descriptors commonly applied to sonar imagery. In these experiments, combinations of the BRISK, FREAK, SIFT and SURF algorithms are parameter-tuned and evaluated to determine the optimal detector-descriptor configurations for bathymetric feature ex-traction. Feature matching is reported using a bathymetric map of Delaware Bay [1] and a semi-synthetic sonar image dataset. Feature extraction and matching performance is evaluated using precision, recall, and the number of true positive matches be-tween features detected in the semi-synthetic sonar imagery and features detected in the full bathymetric map. SIFT was found to be a robust bathymetric feature detector. For the conditions used in this research, the SIFT detector combined with the BRISK descriptor was shown to be the best detector-descriptor combination for bathymetric feature extraction. Matching SIFT_BRISK features achieved perfect precision and perfect recall on 53 correct feature matches.