A System for Acquisition and Modelling of Ice-Hockey Stick Shape Deformation from Player Shot Videos

Kaustubha Mendhurwar, Gaurav Handa, Leixiao Zhu, Sudhir P. Mudur, Etienne Beauchesne, Marc LeVangie, Aiden Hallihan, Abbas Javadtalab, Tiberiu PopaWorkshops @ CVPR 2020 · DOI

Abstract

This paper presents a portable, low-cost stereo camera system capable of automatically reconstructing the 3D deformation of an ice-hockey stick during high-speed shots. By combining a medial-axis constraint with physics-based Finite Element Modeling (FEM), the system accurately captures the "flex" that is critical to shot performance, even in the presence of noise and player occlusions. Validation against professional motion capture systems demonstrates that the system achieves high accuracy, with bending angle errors between 1.85 and 3.5 degrees.

1. Introduction

In ice hockey, the "flex" or bend of the stick is a primary driver of puck speed and accuracy. Elite players use the stick like a spring, storing elastic energy during the loading phase and releasing it rapidly. However, measuring this deformation in a real-world setting is difficult due to the high speed of shots (requiring >250 fps), the thin geometry of the stick, and frequent occlusions by the player's hands and equipment.

Existing solutions often rely on expensive, non-portable Motion Capture (MOCAP) systems that require physical markers. This study proposes a markerless, automated pipeline using high-speed stereo cameras to reconstruct the 3D shape of the stick during a shot.

2. Methodology

The system utilizes two synchronized FLIR Grasshopper3 high-speed cameras capturing at 275 fps. The reconstruction pipeline involves several key stages to handle noise and occlusion.

2.1. Data Pruning and Medial Axis Extraction

To isolate the stick, the system uses temporal information to remove static background elements and proximity-based pruning to remove the player’s body. Because the resulting point cloud is often noisy, the system calculates the medial axis (the central skeleton) of the stick and fits a 3D quadratic curve to it using a RANSAC strategy. This effectively "denoises" the stick's geometry.

2.2. Physics-Based Modeling

To ensure the 3D shape is physically plausible, a co-rotational Finite Element Method (FEM) is employed. A pre-scanned 3D template of the stick is tetrahedralized and constrained to match the observed medial axis curve. The FEM model then "bends" the template to match the data while respecting the material properties of the stick, ensuring topological consistency throughout the shot.

Overview of the 3D reconstruction pipeline: (a) raw stereo frames, (b) initial point cloud, (c) medial axis extraction, (d) curve fitting, (e) FEM-based deformation, and (f) detailed view of the final fit.

3. Results and Validation

The system's performance was evaluated against a professional marker-based MOCAP system.

3.1. Accuracy

The system's calculated maximum bending angles were highly accurate, with differences from MOCAP data ranging between 1.85 to 3.5 degrees. The co-rotational FEM approach proved robust for handling the significant bending and rotation characteristic of high-energy slap shots.

Detailed view of the final FEM-based reconstruction (red) overlaid on the initial noisy point cloud (blue), demonstrating the system's ability to resolve geometry despite occlusion.

3.2. Automation and Efficiency

Aside from initial parameter setup, the pipeline is fully automatic. Currently, processing takes approximately 100 seconds per frame, which the authors aim to optimize for real-time coaching applications.

4. Conclusion and Future Work

This research provides a robust method for modeling thin-object deformation in high-speed environments. The primary limitation remains the stick blade, where proximity to the floor and puck creates geometric ambiguity. Future work will focus on incorporating image-based detection to augment stereo point clouds and developing specific algorithms to separate the blade from the surrounding environment.

Summary updated

BibTeX
@inproceedings{mendhurwar2020system,
  author = {Kaustubha Mendhurwar and Gaurav Handa and Leixiao Zhu and Sudhir P. Mudur and Etienne Beauchesne and Marc LeVangie and Aiden Hallihan and Abbas Javadtalab and Tiberiu Popa},
  title = {{A System for Acquisition and Modelling of Ice-Hockey Stick Shape Deformation from Player Shot Videos}},
  booktitle = {Workshops @ CVPR},
  year = {2020},
  doi = {10.1109/cvprw50498.2020.00453},
}
Figure