Surfing Pose Estimation & Maneuver Classification

Surfing analytics with robust pose tracking (YOLOPose/RTMO) and sequence-based maneuver classification (TCN/PosFormer).

Categories

Tech Used

Problem

Surfers and coaches need fast, objective feedback on maneuver execution and technique without manual video review, even in challenging surf conditions (spray, occlusions, moving cameras).

Approach

• Implemented multi-backend pose estimation with YOLOPose and RTMO for resilient keypoint tracking in real surf footage
• Stabilized joint trajectories using temporal filtering (Kalman/Butterworth) to reduce jitter and improve downstream recognition
• Built a temporal maneuver classifier using TCN and PosFormer to model motion dynamics across frames and detect maneuvers (e.g., bottom turn, top turn, cutback, snap)
• Designed a feature pipeline from pose sequences (joint angles, stance geometry, rotation cues, velocity/acceleration) to support robust classification and segmentation
• Deployed real-time inference with streaming video input and low-latency feedback via WebSockets and a backend API

Results

• Stable pose tracking under occlusions, water spray, and rapid camera motion
• Accurate maneuver segmentation and classification using TCN/PosFormer temporal modeling
• Real-time coaching feedback loop ready for integration into training apps

Demo Videos