To address the key challenges in 3D motion estimation of sonar sensors, this paper proposes a novel sonar odometry approach. The proposed method establishes a complete 3D motion estimation pipeline for sonar, incorporating two-view acoustic bundle adjustment, 3D-2D acoustic Perspective-n-Point (PnP) solving, and global bundle adjustment optimization. Without relying on planar assumptions or pre-defined artificial acoustic landmarks, the proposed method introduces a graph-based acoustic structure-from-motion technique to overcome the failure of analytical closed-form solutions caused by the nonlinear projection model. Additionally, a motion direction prior is employed to effectively resolve the symmetric dual-pose ambiguity resulting from elevation angle uncertainty in sonar measurements. The integration of global bundle adjustment allows the system to fully utilize historical observations, significantly improving pose estimation accuracy. The effectiveness of the proposed method are validated on datasets generated in a self-developed simulation environment, covering both 2D planar scenes and randomly generated 3D environments.