Tools/AI tools/Human Pose Estimator
PIXELOKAY · MOTION INTELLIGENCE ✦ AI Studio

Human Pose Estimator

Detect body joints, inspect skeleton overlays, trace movement and export actual pose landmarks frame by frame. AI inference happens in your browser, not on our servers.

✓ 33 body landmarks✓ Motion trails & joint angles✓ CSV + JSON export
YOUR POSE ANALYSIS WORKSPACE

Human Pose Estimator AI / Beta

Upload media, configure sampling, analyze, and inspect your results.

01
Preview & skeleton overlayNo media selected
Model not loaded
◎
Drop an image or video here

PNG, JPG, WebP or browser-decodable video

Files stay on this device · up to 180 MB
0.00s / 0.00s
No poses analyzed
02
Analyze pose dataModel downloads only when you click Analyze
Local AI

Choose a media file to estimate the analysis workload.

Sample data is clearly labeled synthetic. Real pose results require the free MediaPipe model and an initial internet connection. No media file is transmitted by this app.

03
Frame-by-frame pose resultsNever infer unseen joints or fabricate missing poses
0 frames
Analyzed0frames
Detected0%with a pose
Peak people0per frame

Joint measurements

Select an analyzed frame
JointXYZVisible
Analysis results will appear here.

X/Y are normalized image coordinates, not physical distances. Z is model-relative, not centimeters. Estimated world coordinates are included in the JSON and optional world CSV where available.

Choose a source image or video to begin.

From footage to structured movement data.

◎

Movement analysis

Track wrists, knees and other joints to visualize motion through a sampled video.

⌁

Skeleton overlays

Render joints, limb connections and visibility-filtered labels on top of a frame.

↗

Open data

Export normalized landmarks, optional model-estimated world coordinates and project JSON without an account.

Frequently asked questions

Is this genuine AI pose estimation?

Yes, when you click Analyze the page loads Google's free MediaPipe Pose Landmarker model and performs inference locally. The sample button uses synthetic demonstration values, which are labeled explicitly.

Does my video leave my browser?

The app does not upload media. The MediaPipe JavaScript, WebAssembly runtime and model are initially fetched from external hosts. MediaPipe may collect library usage metrics; see their privacy information. Self-host those files to eliminate the runtime CDN dependency.

Why are some joints missing?

Hidden, occluded, fast-moving or out-of-frame joints may not be predicted reliably. The visibility cutoff controls which predictions are drawn, but the raw output remains available in exported data.

Is it suitable for clinical measurements or gait diagnosis?

No. This tool provides approximate computer-vision estimates, not validated clinical or biomechanical assessments. Angle calculations use 2D image coordinates and change with camera perspective.

Will WebM preserve my video audio and original frame rate?

No. Video export is silent and includes only analyzed sample frames. It uses real-time browser Canvas recording, so timing and codec support vary. CSV and JSON are the most faithful exports for frame-by-frame measurements.

Using Human Pose Estimator

Estimate pose landmarks from a photo or sampled video, then review trajectories and approximate angles. Treat the output as visual data for exploration, not calibrated measurement or identity tracking.

  1. Choose a photo or video, or inspect the clearly labeled synthetic sample, then set the model and sampling range.
  2. Review the model setup details and explicitly load compatible model assets before requesting pose analysis.
  3. Inspect the landmark overlay, missing detections and approximate angles; export reviewed data or a selected visual result.

Example: For a two-second side-view exercise clip, inspect the reported hip and knee landmarks across sampled frames, note where detections disappear, and export the reviewed angle estimates for further analysis.

Limits and privacy: The tool accepts up to four people and 420 sampled frames, with video duration capped at 120 seconds. Model/runtime assets are fetched from external hosts after explicit setup and may disclose the visitor IP; actual inference remains unverified. World coordinates and 2D angles are estimates, person ordering is not persistent identity, and missing detections remain gaps. Browser-recorded WebM is silent and may vary by browser.

Related tools

LOCAL AI SETUP

Set up AI on this device

First, a model needs to be loaded on your device

Human Pose Estimator uses the free MediaPipe pose model and WebAssembly. First setup downloads official model/runtime files. Your photos and video stay in this browser; optional MediaPipe usage metrics may be sent by its SDK.

  1. 1. Check device
  2. 2. Download / cache
  3. 3. Initialize pose detector
  4. 4. Analyze media
Choose a MediaPipe pose model. Download model assets explicitly; images and videos are analyzed locally.

Device & model resources

Measurements may be approximate
Device RAM / model memoryChecking device information…

Exact RAM and GPU memory are not available. Pose Landmarker uses WebAssembly; WebGPU is not required.

Browser website storageChecking storage…

Site-wide estimate; includes other PixelOkay tools and is not RAM.

ONE ACTION · ALL RECOGNIZED AI MODEL DOWNLOADS

Free AI model storage across PixelOkay

Scan this exact website origin for recognized WebLLM, Transformers.js, MediaPipe pose, and other safely identifiable model caches (including those from other PixelOkay tools). This does not clear general website data or browser databases.

Run a scan to see how many model caches can be cleared together.

Browser-native models and unknown IndexedDB/OPFS or private runtime caches may remain. Unknown application/user storage will not be deleted. Other running tools might redownload caches; close their tabs first.

Having trouble downloading or loading AI?

If a model fails, retry, check your network, or refresh model caches. Incompatible worker environments may use an explicit main-thread fallback; it can briefly block the interface.

Only model caches are targeted. Other PixelOkay tools may need to redownload their models. Never clear all website data without exporting important chats and projects first.

Select Check compatibility to inspect whether WebAssembly and workers are available.
Technical details and compatibility fallback

Pose runtime: MediaPipe Tasks Vision 0.10.35 · official Lite/Full pose .task files. Inference takes place locally, with an optional main-thread compatibility mode.

CROSS-TOOL MODEL CACHE CLEANUP

Delete recognized AI model downloads?

Only independently recognizable model caches are targeted. Other PixelOkay tools may need to redownload models. Saved projects, mapping rules, files, and browser databases are not selected, but unknown stores are not inspected or removed.

Close other AI tabs before proceeding. Browser-managed, IndexedDB, and OPFS model files may remain.