Person Detection with Distance, Gesture, and Player Status Prediction

Developed a YOLO-based person detection model for tennis court videos that predicts both person bounding boxes and additional person-level attributes in a single perception model.

For each detected person, the model predicts camera-to-person distance, hand-raise gesture status, and player role status, including whether the person is the active tennis player or a bystander. This allows the tennis perception system to understand not only where people are in the image, but also their distance, behavior, and relevance to the match.

Built the dataset preparation workflow from real tennis video frames using automated person and depth labeling, ReID-based active-player tracking, dataset statistics analysis, distance outlier filtering, and YOLO-format annotation generation.

I trained and validated the custom person-attribute detection task, then implemented inference scripts to visualize bounding boxes, distance predictions, gesture status, and active-player classification on tennis video streams.

The model was also prepared for edge deployment through ONNX export, TorchScript export, calibration image generation, INT8 post-training quantization, and inference comparison on RDK X5 hardware.

This model is one component of a complete tennis scene understanding system, where person detections are later combined with court pose estimation and ball tracking to estimate player positions in real-world 3D court coordinates.