Live Pose Detection

Real-time human pose estimation system for fitness and healthcare. Captures 33 full-body keypoints via webcam, computes joint angles, counts exercise reps, scores form quality, analyzes posture, and exports session data — all through a desktop UI.

Features

• Real-time pose tracking — MediaPipe PoseLandmarker, 33-keypoint full-body, 30+ FPS on CPU
• Exercise rep counting — Squat and push-up detection with configurable angle thresholds
• Form scoring — Green/yellow/red skeleton overlay based on joint angle quality
• Posture analysis — Detects forward head, uneven shoulders, rounded upper back, neck tilt
• Live angle timeline — Real-time matplotlib chart of joint angles over the last 300 frames
• Form guide overlay — Semi-transparent reference skeleton for pose comparison
• Video file input — Load pre-recorded .mp4 / .avi / .mov files in addition to live camera
• Session recording — Annotated MP4 video + raw keypoint data (CSV/JSON) per session
• Multi-person support — Up to 3 people tracked with proximity-based ID assignment
• Desktop UI — PySide6 with live camera feed, skeleton overlay, and tabbed stats panel

Tech Stack

Component	Technology
Pose Engine	MediaPipe PoseLandmarker (BlazePose)
GUI	PySide6
Computer Vision	OpenCV
Math	NumPy, SciPy
Analytics	Matplotlib, Pandas
Package Manager	uv
Testing	pytest

Quick Start

Requires Python 3.11+ and uv.

git clone git@github.com:abuzarai/Live-Pose-Detection.git
cd Live-Pose-Detection
make install
make run

CLI Options

make help

--camera 2        Use external camera
--video demo.mp4  Analyze a recorded video
--exercise Posture Healthcare posture analysis mode
--show-guide      Display reference skeleton overlay

Controls

Button	Action
Start Camera	Opens webcam feed
Open Video	Load a video file for analysis
Exercise dropdown	Switch between Free / Squat / Push-up / Posture modes
Record	Save annotated MP4 + CSV/JSON to data/sessions/
Screenshot	Capture current frame as PNG

Project Structure

live_pose_detection/
├── main.py                    # CLI entry point
├── core/
│   ├── detector.py            # MediaPipe PoseLandmarker wrapper
│   ├── calculator.py          # Joint angle + posture calculations
│   └── tracker.py             # Multi-person ID + landmark smoothing
├── ui/
│   ├── main_window.py         # PySide6 window (camera, controls, tabbed panels)
│   ├── overlay.py             # Keypoint/skeleton drawing with form colors
│   └── panels.py              # Angle list, stats, posture feedback, timeline chart
├── features/
│   ├── exercise.py            # Squat/push-up detectors with rep counting
│   ├── posture.py             # Forward head, shoulder slope, spine curve analysis
│   └── recorder.py            # Session recording (MP4 + CSV/JSON export)
├── utils/
│   └── config.py              # JSON settings loader/saver
└── data/
    ├── config.json            # Persistent settings
    ├── models/                # MediaPipe model files
    └── sessions/              # Recorded session output

Testing

make test

15 tests covering angle calculations, exercise detection logic, and config persistence.

License

MIT