Abstract: Three-dimensional motion estimation serves as a crucial technology for autonomous underwater vehicles to attain high-precision navigation and positioning. This paper proposes a three-dimensional visual odometry framework based on forward-looking sonar, which consists of three main modules: feature extraction, elevation map restoration, and three-dimensional motion estimation. In the feature extraction module, a sliding window algorithm based on Mahalanobis distance is combined with the local binary fitting level set algorithm to achieve fine-grained segmentation of the target contour. Additionally, a target-shadow pair is constructed based on the gray-scale distribution characteristics and gradient properties of the target. In the elevation map restoration module, a nonlinear equation system is constructed based on the Lambert diffuse reflection model, and the elevation of the target internal points is iteratively solved. In the three-dimensional motion estimation module, the coherent point drift point cloud registration method is used to register the elevation maps and solve the three-dimensional motion parameters. Results from simulations and field measurements demonstrate that, when compared with other algorithms, the proposed method outperforms them in evaluation metrics such as average error, cumulative error, and root mean square error. Evidently, it significantly enhanced the accuracy and robustness of motion estimation.