>embedforge
← Back to projects

Case study

AirDrawingAI

Hand-tracking drawing app that turns finger movement into on-screen strokes — no stylus, no touch.

PythonOpenCVMediaPipeComputer Vision
Tracking latency
~35ms
Frame rate (CPU only)
24-30 FPS
Gesture accuracy
~96% in controlled lighting

Problem

Touchless input for whiteboarding and accessibility use cases needed sub-50ms hand landmark tracking robust enough to run on a mid-range webcam and laptop CPU, without a GPU.

Stack

  • Python
  • OpenCV
  • MediaPipe Hands
  • NumPy
  • Tkinter UI overlay

Architecture

AirDrawingAI system architecture: webcam feed, MediaPipe hand tracking with 21-point landmark detection, gesture classifier reading fingertip position and pinch state, OpenCV canvas drawing strokes at the fingertip, and display output compositing the overlay on video.

Outcome

Runs at consistent frame rate on integrated graphics with a lightweight Kalman-style smoothing filter to remove landmark jitter, making single-pixel strokes usable for fine control.

View on GitHub ↗