Aircraft attitude deviation during the approach and landing phase is an important indicator for operational risk identification and active warning. To address the difficulty of traditional attitude sensing methods in forming independent and non-contact active detection at the ground side, a ground vision-based attitude estimation method integrating structure-constrained sparse keypoint detection and globally optimal PnP solving is proposed. A multiscale monocular video dataset is constructed using high-fidelity flight simulation and three-dimensional rendering. Based on YOLO11n-pose, a dynamic feature fusion module and a graph-convolutional pose refinement module are introduced for keypoint recognition. An integrated solution chain from two-dimensional keypoints to attitude angles in the navigation coordinate system is then established by combining adaptive RANSAC, globally optimal SQPnP, multi-coordinate-system transformation, and causal bilateral temporal filtering. Experimental results show that the proposed method achieves higher keypoint detection accuracy and attitude estimation accuracy than the baseline model. The root mean square errors of roll, pitch, and yaw are 0.137°, 0.963°, and 0.623°, respectively, and the average total processing time per frame is about 19 ms. The proposed method provides reliable technical support for independent non-cooperative ground-side attitude monitoring in aircraft approach and landing scenarios.