Grid-Guided Tennis Court Keypoint Detection

Developed a grid-guided tennis court keypoint detection model for identifying court line intersection points from camera images.

The model combines the original RGB image with a constant synthetic grid image as an additional input channel group. This gives the model spatial reference without drawing grid lines over the real image, preserving the original visual content.

For deployment, the constant grid is embedded directly inside the exported ONNX graph, allowing the model to accept a standard RGB image while internally adding the grid representation during inference.

The detected court keypoints are used for homography and camera pose estimation, enabling downstream player mapping, 3D ball trajectory analysis, automated line calling, and full tennis scene understanding.