Geometric Modeling of Strawberry Flower Spatial Orientation via Multi-Keypoint Plane Fitting and Orthogonal Projection of the Pistil Center onto the Corolla Plane

Robotic pollination of greenhouse strawberries requires knowing not only where a flower is, but also the direction in which it opens. Existing vision approaches, however, mostly deliver 2D bounding boxes or single-point depth, which is insufficient for planning the end-effector’s approach pose. This paper presents a keypoint-driven geometric method for estimating strawberry flower orientation in 3D. A lightweight pose network, YOLOv8-Pose-CA-RepVGG, was constructed by embedding RepVGG re-parameterization and Coordinate Attention into the YOLOv8-Pose backbone to locate petal-edge keypoints and the pistil center. These keypoints are back-projected into the camera coordinate system with aligned RGB-D data, and the corolla reference plane is reconstructed through least-squares fitting based on singular value decomposition. The plane normal is then refined by the orthogonal projection of the pistil center, yielding a unit vector that points from the corolla toward the pollination target. On the main greenhouse dataset of 873 images, the model achieved 80.40% mAP@0.5 and 67.20% mAP@0.5:0.95, improvements of 4.10 and 3.00 percentage points over YOLOv8-Pose, respectively. On the specified NVIDIA GeForce RTX 3050 Ti Laptop GPU, it achieved an inference throughput of 267.65 FPS at a 640 × 640 input resolution and a batch size of 16. In the independent normal-vector evaluation set, valid orientation outputs were obtained for 245 of 300 samples (81.67%). Both valid model outputs and evaluable reference normals were available for 183 of the 300 samples (61.00%; 49 Chunri, 68 Hongyan, and 66 JW8). On this reference-evaluable subset, conditional Accuracy@60° was 91.84% for Chunri, 92.65% for Hongyan, and 90.91% for JW8. These results characterize keypoint detection and geometric normal-vector estimation under the evaluated greenhouse conditions; validation of pollination performance on a robot platform remains necessary.