YMori

Posted on Feb 25 • Edited on Mar 1 • Originally published at zenn.dev

3D Skeleton Detection from Baseball Motion Capture Data with Driveline C3D

#python #baseball #datascience #computervision

License: Motion capture data from Driveline OpenBiomechanics Project under CC BY-NC-SA 4.0.
Citation: Wasserberger KW, Brady AC, Besky DM, Jones BR, Boddy KJ. The OpenBiomechanics Project: The open source initiative for anonymized, elite-level athletic motion capture data. (2022).
License: https://creativecommons.org/licenses/by-nc-sa/4.0/
Derivative works (graphs, GIFs) in this article follow the same license.

What I Built

I visualized baseball pitching and hitting in 3D, extracted joint kinematics, and explored the relationship between body mechanics and pitch velocity.

Tools used:

Driveline OpenBiomechanics Project (OBP) — elite-level motion capture C3D data (100 pitchers + 98 hitters)
ezc3d — C3D file I/O library (GitHub, MIT License)
matplotlib — 3D visualization and animation

→ GitHub: https://github.com/yasumorishima/baseball-cv

My Connection to ezc3d

I contributed a bug fix to ezc3d via PR #384. Using a library I contributed to for this analysis felt like a natural progression.

Step 1: 3D Skeleton Visualization from C3D

C3D files contain 3D coordinates of body markers captured by motion capture systems.

Pitching: 45 markers, 360Hz, ~726 frames
Hitting: 55 markers (45 body + 10 bat), 360Hz, ~804 frames

import ezc3d

c3d = ezc3d.c3d("pitching_sample.c3d")
points = c3d["data"]["points"]  # shape: (4, n_markers, n_frames)
labels = c3d["parameters"]["POINT"]["LABELS"]["value"]

Pitching Skeleton Animation

45 markers connected as a skeleton. The full wind-up to release motion is visible.

Hitting Skeleton Animation

55 markers with bat markers shown in red.

Step 2: Video-Based Skeleton Detection with MediaPipe

Beyond C3D data, Google's MediaPipe Pose can detect 33 body landmarks from regular video.

import mediapipe as mp

mp_pose = mp.solutions.pose
pose = mp_pose.Pose(
    static_image_mode=False,
    min_detection_confidence=0.5,
    min_tracking_confidence=0.5
)

This article analyzed C3D data from motion capture, but MediaPipe's advantage is that it can work from regular smartphone video — no specialized equipment needed.

Step 3: Joint Angle & Angular Velocity Extraction

From the skeleton coordinates, I computed joint angles as time series.

Pitching Results

Joint Angle	Min	Max
Elbow Flexion	50.5°	156.7°
Shoulder Abduction	4.6°	117.7°
Trunk Rotation	0°	58°
Knee Flexion	99.1°	163.8°

Angular Velocity Time Series

This plot shows angular velocity (degrees/sec) per frame, revealing which joints move fastest at each phase of the pitching motion.

Step 4: Skeleton Features × Pitch Velocity Correlation

Driveline OBP C3D filenames encode pitch velocity (e.g., ..._809.c3d → 80.9 mph).

I analyzed 16 pitchers to find correlations between skeleton-derived features and velocity.

Correlation Results

Feature	r	p-value
Peak Trunk Angular Velocity	0.119	0.673
Peak Elbow Angular Velocity	0.094	0.739
Peak Shoulder Abduction	0.180	0.520
Trunk Rotation Range	0.425	0.114

With only 16 samples, none reached statistical significance. However, trunk rotation range of motion showed the strongest positive correlation (r=0.425) with pitch velocity.

This suggests that "how far a pitcher can rotate their trunk" may contribute to velocity. A larger sample size would likely clarify the relationship.

Summary

Visualized 3D pitching and hitting skeletons from Driveline OBP C3D data using ezc3d
Extracted joint angle and angular velocity time series
Trunk rotation range showed the strongest correlation with pitch velocity (r=0.425)
MediaPipe can detect skeletons from regular video (this article used motion capture C3D data)