High-precision self-supervised underwater visual odometry via unified image enhancement and coupled cycle consistency

Abstract

The purpose of this paper is to propose a high-precision self-supervised monocular visual odometry (VO) framework for navigation in complex underwater environments. An uncertainty-aware underwater image enhancement module is integrated into a self-supervised monocular VO framework. By embedding an uncertainty map estimation branch and fusing features extracted from both pre-enhanced and post-enhanced images, image enhancement and pose estimation are jointly optimized. Furthermore, a coupled cycle consistency loss is introduced to exploit the intrinsic relationship between depth information and underwater transmission maps, thereby improving depth estimation accuracy through mutual supervision. Experimental results obtained on the publicly available FLSea-VI data set indicate that the proposed method achieves higher depth estimation and pose estimation accuracy than existing state-of-the-art monocular VO approaches. The results further suggest that the proposed framework is well suited to challenging underwater environments. The proposed framework may offer incremental safety benefits for diving operations and autonomous navigation. Furthermore, by lowering technical barriers, it could facilitate marine exploration for smaller research teams, thereby encouraging broader participation in marine environmental monitoring and related endeavors. A unified self-supervised underwater visual odometry framework is presented in which image enhancement, depth estimation and pose estimation are jointly optimized. An uncertainty-aware enhancement strategy and a coupled cycle consistency mechanism are introduced to address underwater image degradation and the ill-posed nature of monocular depth estimation.