Abstract: Accurate perception of the heading of surrounding unmanned surface vessels is a critical foundation for the safe and stable operation of unmanned surface vessels during coordinated operations and autonomous navigation missions. This paper designs a monocular vision unmanned vessel course perception architecture based on spatial-orientation normalization by actively recognizing the navigation status of surrounding unmanned vessels through a single camera. It utilizes the YOLOv11-pose model to extract the coordinates of key points of the unmanned vessel structure, and introduces an improved coordinate attention mechanism to address the difficulties in locating the geometric structures of unmanned vessels at the bow and stern in complex environments. Additionally, a key point feature learning strategy based on spatial-orientation normalization is designed, with the bow-to-stern vector as the main axis, to construct a local coordinate system and achieve scale invariance of unmanned vessel translation and rotation. Finally, a multi-head attention-enhanced perception network is used to model and recognize the normalized coordinate points of the unmanned vessel structure in the image, and to identify the course of the unmanned vessel. In experiments, we extract the geometric structure coordinates of unmanned vessels from real unmanned vessel navigation images and identify the course of the unmanned vessels. The key point detection model adopted achieves the best results in indicators assessment such as Recall and mAP@50, and the accuracy of course identification is improved to nearly 90%, with an accuracy improvement of 11% compared to the basic framework. The experimental results demonstrate the effectiveness of the monocular vision course perception architecture proposed in this paper in the course perception task.