A minimal browser-based kettlebell swing coach using on-device pose inference, personalized calibration, spatial Gaussian confidence fields, and live layered 3D anatomy.

• Runs MediaPipe Pose Landmarker in the browser against a webcam stream.
• Uses 33 pose landmarks plus world-coordinate landmarks for hinge, knee, torso, shoulder, and depth-path analysis.
• Calibrates to the user's standing posture before scoring reps.
• Tracks swing phase, rep count, hinge-to-knee ratio, lockout, shoulder lift, spine stack, depth travel, camera quality, and confidence.
• Transforms the mirrored camera feed into projected body, muscle, skeleton, and Gaussian correction layers.
• Shows a Three.js anatomical rig with independently toggled body, muscle, skeleton, and Gaussian field layers.
• Renders procedural muscle volumes for glutes, hamstrings, quads, calves, spinal erectors, core, lats, deltoids, and forearms with intensity tied to the current swing metrics.

npm install
npm run dev

Open the local Vite URL, start the camera, then calibrate while standing tall in frame before swinging.

The MVP is intentionally browser-first and on-device:

• MediaPipe Pose Landmarker outputs normalized image landmarks and 3D world landmarks in meters, and supports frame-by-frame video inference in JavaScript. https://ai.google.dev/edge/mediapipe/solutions/vision/pose_landmarker/web_js
• BlazePose GHUM was designed for real-time, on-device 3D pose estimation from a single RGB image, making it a practical base for live coaching. https://arxiv.org/abs/2206.11678
• Depth Anything V2 is a stronger route for future dense pixel-depth fusion, but the current app uses pose-world depth and calibrated wrist/landmark travel so it stays performant in a live browser session. https://github.com/DepthAnything/Depth-Anything-V2
• Kettlebell swing research supports treating swing styles as biomechanically distinct; this app targets the hardstyle/Russian hip-hinge swing pattern, not every valid swing variation. https://pubmed.ncbi.nlm.nih.gov/28593086/
• Hamstring EMG research found hip-hinge swings produced greater hamstring activity than squat and double-knee-extension swing styles, which is why the coach penalizes squat-dominant reps for this target pattern. https://pubmed.ncbi.nlm.nih.gov/28930870/

The coach does not claim clinical-grade motion capture. It improves reliability by combining:

• calibration against the user's upright hip, knee, torso, shoulder-width, and torso-length profile
• side-view camera quality checks
• landmark visibility and jitter weighting
• 3D hip/knee angles from world landmarks
• 2D screen-space checks for bell/hand height and camera framing
• depth travel from wrist world coordinates
• Gaussian overlays to show uncertainty and local error risk
• projected 2D anatomy layers and procedural 3D anatomy layers anchored to tracked pose landmarks

For higher precision, the next engineering step is an optional dense-depth worker using Depth Anything V2 or a WebGPU/Transformers.js depth model, fused with pose landmarks and a camera calibration routine.