Robust Underwater Object Detection and Relative Navigation Using Simulation-Aided Sonar Image Translation

Abstract

: Robust underwater object detection and relative navigation remain challenging because of the severe degradation of sonar imagery caused by acoustic noise and artifacts. Although imaging sonar provides a viable sensing modality in GPS-denied underwater environments, its effectiveness is limited by low signal-to-noise ratios, multipath interference, and domain-dependent intensity distortions that hinder reliable feature extraction. This paper presents a simulation-aided unpaired image translation framework for enhancing sonar-based perception and navigation. The proposed method uses an unpaired image translation technique for mapping from noisy real sonar images to a noise-suppressed target domain constructed from simulated sonar images with controlled noise characteristics, without requiring paired training data. The translated images were processed using an object detector, and the resulting detections were incorporated as landmark observations within a landmark-based relative navigation framework. The experimental results with real sonar data show that the proposed approach substantially reduces false detections, leading to significantly improved detection precision and trajectory estimation accuracy compared to applying detection directly on raw sonar images. These results confirmed that simulation-aided perception can effectively bridge the domain gap in sonar-based underwater navigation without requiring paired training data.