Project · L05 Intelligence
Real-Time Hand Tracking
A hand, tracked from a webcam in real time, conducting a field of particles — the project behind the particles on this site.
- Role
- Full-stack developer
- Period
- Nov 2025 — Dec 2025
- Stack
-
- C++20
- MediaPipe
- OpenCV
- OpenGL
- SDL2
- Python
Context
I wanted to understand the full loop of a real-time vision system — capture, inference, smoothing and rendering — inside a single frame budget.
What I built
- Low-latency capture that always processes the latest camera frame.
- MediaPipe Hands inference with an optional GPU delegate and a CPU fallback.
- One Euro filtering and confidence-aware pose stabilisation to remove jitter without adding lag.
- Gesture-driven effects on a modular effect system, with particle pooling for stable frame times.
- An on-screen overlay showing FPS, inference time, render time, the detected gesture and its confidence.
Stack
Prototyped with Python, MediaPipe and OpenCV. The repository implementation uses C++20, MediaPipe Tasks, OpenCV, SDL2, OpenGL 3.3, GLEW and CMake.
On this site
The particle field you scrolled through is a descendant of this project. The hand in layer L05 is built on MediaPipe’s 21-landmark hand topology — the same skeleton the tracker follows.