Abstract: Unmanned surface vehicle(USV) requires accurate pose estimation of the target during dynamic docking operations. However, large observation angles, motion blur, and occlusion caused by target motion can lead to feature degradation, making traditional planar markers inadequate for reliable omnidirectional detection. To address this problem, this paper proposes a vision-based pose detection method using a stereo marker. Firstly, a four-faced stereo marker composed of a multi-ring nested structure is designed, which enables full-attitude detection by utilizing circular projection features and stereo distribution properties. Secondly, an arc segment matching method based on ring projection and concentricity constraints is introduced, allowing stable recognition even when marker features degrade or are partially missing. Next, a target localization method based on an imaging model is proposed, which corrects circular features through affine transformation and optimizes localization results using the geometric properties of a regular quadrangular prism. Finally, continuous attitude estimation is achieved by incorporating the temporal trend of target motion. Simulation results show that the maximum position error is below 0.8 m and the maximum heading error is less than 5° within a range of 30 m, verifying the accuracy of the method. Real-world ship experiments demonstrate that the position error is below 1.2 m and the maximum heading error is under 6.2° in static tests, and the position error at the final stage is below 0.16 m and 85% of the heading errors are within 5° in dynamic docking tests, further confirming the feasibility and practicality of the proposed method.