Hi, I'm Michael.

Built a Python pipeline that syncs phone videos of judo matches, filmed from different viewpoints, to within one frame at 30 fps. It runs 2D pose estimation (Pose2Sim, RTMPose) and cross-correlates filtered keypoint velocities (r = 0.98), detecting the opening bow from where the head drops below the shoulder line, which replaces manual timestamping. Showed that the existing YOLO-based match-start detector missed the bow at its 1 Hz sampling rate, which motivated the pose-based approach.