Pose Estimation Studio
Detect human body landmarks from a live webcam or an uploaded image, then study how skeleton tracking supports fitness analysis, posture understanding, and human activity recognition.
Webcam preview
Click Start Camera to grant permission and see real-time pose keypoints.
Camera controls
Start webcam and track poses
The demo only starts the webcam when you click “Start Camera”.
Camera not started
Privacy note
- Webcam frames are processed locally in your browser.
- No video is uploaded to a server by this demo.
- The camera does not start automatically. Click “Start Camera” to request permission.
- You can stop the camera at any time.
Status
Model ready
Loading model...
Detected people
0
Keypoints shown
0
Camera
Camera not started
Visual settings
Keypoint display
Use toggles to reduce clutter, and adjust confidence to hide uncertain keypoints.
Toggles
Higher threshold hides more low-confidence keypoints. Lower threshold shows more, but can include noise.
Interpretation
What the overlay means
Loading the model...
Reminder: pose estimation predicts body geometry. It does not identify who the person is.
Keypoints
Coordinates & confidence
X and Y are in the video frame coordinate system. Confidence indicates how certain the model is.
Body region summary
How much is visible
Face
0
Upper
0
Lower
0
This summary counts keypoints above the current confidence threshold.
Movement analysis (educational)
From keypoints to motion
This demo builds movement intuition from pose keypoints. It is not a trained medical or performance classifier.
Arms raised
—
Wrist keypoints above shoulder level.
Squat / bend cue
—
Knees lower than hips (approx rule).
One wrist above shoulder
—
A simple educational rule that helps illustrate how pose tracking changes over time.
Body center movement
—
Approximate displacement of the hip center across recent frames (educational estimate).
In real applications, movement analysis usually requires more robust models, testing, and careful safety validation.
Body keypoint map
Common landmarks the model predicts
Pose detection predicts important points on the body. It does not identify who the person is.
Nose
A reference point near the center of the face.
Left eye
One of the face landmarks used to estimate head geometry.
Right eye
One of the face landmarks used to estimate head geometry.
Left ear
Helps stabilize pose when the face is partially visible.
Right ear
Helps stabilize pose when the face is partially visible.
Left shoulder
Upper-body landmark used for arm angle estimation.
Right shoulder
Upper-body landmark used for arm angle estimation.
Left elbow
Middle arm joint that helps track bending and reach.
Right elbow
Middle arm joint that helps track bending and reach.
Left wrist
Hand joint used for movement and pointing gestures.
Right wrist
Hand joint used for movement and pointing gestures.
Left hip
Core landmark for posture and lower-body movement cues.
Right hip
Core landmark for posture and lower-body movement cues.
Left knee
Lower-body landmark for squat and step movement cues.
Right knee
Lower-body landmark for squat and step movement cues.
Left ankle
Foot landmark used with knees/hips to estimate leg position.
Right ankle
Foot landmark used with knees/hips to estimate leg position.
Step-by-step pipeline
How pose detection becomes movement cues
This pipeline explains the flow of real-time inference and thresholding.
Webcam frame input
A webcam provides many frames per second for real-time inference.
Frame preprocessing
The browser prepares the frame for the model (resize/format as needed).
Body region detection
The model estimates where a person is in the frame.
Keypoint prediction
It predicts body landmarks like shoulders, elbows, wrists, hips, knees, and ankles.
Confidence scoring
Each keypoint gets a confidence score showing how certain the model is.
Skeleton connections
When keypoints are confident enough, the demo draws lines between related joints.
Coordinates displayed
Coordinates and confidence values update as the video changes.
Movement analysis
Movement can be estimated by tracking keypoints over time. Context still matters.
How pose estimation works
Keypoints, not identities
Pose estimation is a computer vision task where an AI model predicts keypoints on a human body, such as shoulders, elbows, wrists, hips, knees, and ankles.
It does not identify who a person is. It estimates body geometry and movement patterns for learning purposes.
Keypoints can be connected to draw a skeleton. By tracking changes over time, we can build simple movement analysis.
This demo uses a pre-trained model and runs inference in your browser. It does not learn from your webcam or uploaded images during the demo.
Keypoint coordinate and confidence explained
How to interpret the numbers
X coordinate
Horizontal position of the keypoint in the frame or image.
Y coordinate
Vertical position of the keypoint in the frame or image.
Confidence score
How certain the model is about the keypoint.
Confidence threshold
The minimum confidence needed to draw dots or lines.
Real-time inference means predictions repeat across video frames, while uploaded images let students inspect the same body geometry as a static example.
Movement analysis explanation
How motion cues can be built from keypoints
Arm movement can be tracked by following shoulder, elbow, and wrist points over time.
Squat or jump movement can be estimated from hip, knee, and ankle positions.
Posture can be approximated from shoulder, hip, and body alignment patterns.
Sports motion can be studied by tracking keypoints over time, but this demo is not a medical tool.
Pose detection vs other AI tasks
What each task is trying to do
Image classification
- Predicts a label for the whole image.
- Does not locate body parts.
Object detection
- Detects objects and bounding boxes.
- May detect a person, but not detailed joints.
Pose detection
- Predicts body keypoints.
- Draws a skeleton.
- Supports movement analysis.
Face / hand landmarks
- Focus on specific regions.
- Pose detection focuses on whole-body keypoints.
Real-time AI explanation
Why it feels live
- A webcam provides many frames per second.
- The model processes frames repeatedly.
- The canvas overlay updates with predicted keypoints.
- Lighting, distance from the camera, clothing, occlusion, and device performance affect accuracy and smoothness.
- Browser-based AI can keep webcam frames local.
Applications of pose estimation
Where pose tracking is used
Sports movement analysis
Feedback for form and training drills.
Exercise feedback
Guidance by observing posture changes.
Dance and motion learning
Interactive classroom demonstrations.
Rehabilitation research
Studying movement patterns in controlled settings.
Human-computer interaction
Gesture-based interfaces and controls.
Animation and avatar control
Turn keypoints into visual motion.
These applications should be designed carefully with consent, fairness, safety, and validation.
Limitations and ethics
Use it responsibly
- The model can make mistakes.
- Poor lighting can reduce accuracy.
- Fast movement can cause unstable keypoints.
- Occlusion, loose clothing, unusual poses, or parts outside the frame reduce accuracy.
- Camera angle and distance affect results.
- Pose estimation should not be used alone for medical diagnosis, workplace monitoring, grading, policing, legal decisions, or safety-critical decisions.
- This demo is for learning and teaching only.
Student learning outcomes
What you will learn
- Understand what pose estimation means.
- Understand body keypoints and skeleton connections.
- Understand real-time webcam inference and still-image landmark analysis.
- Understand confidence scores and thresholds.
- Understand how pose tracking supports movement analysis.
- Understand the difference between classification, object detection, landmarks, and pose estimation.
- Understand privacy and ethics in body-tracking AI.
- Learn how pre-trained computer vision models can run in the browser.
Final reminder
Webcam frames and uploaded images are processed locally in your browser. They are not uploaded to a server, and the model is pre-trained.