Project
DJ Pose Estimation
Overview
DJ Pose Estimation is a browser-based, pose-driven music visualizer created in the scope of the course "Visual Analysis of Human Motion". I collaborated on the project with my university colleague David Köppl.
The project addresses a practical gap for entry-level and hobby DJs: audio-reactive visualizers are easy to access, but they usually give performers little direct control over the visuals. Our prototype uses webcam-based arm gestures to steer a real-time particle system, so performers can influence the visual output without additional hardware or a dedicated video artist.
The application combines MediaPipe Pose, TensorFlow.js, React, TypeScript, Three.js, and WebGL. We trained a lightweight five-pose classifier on custom videos, converted the model for browser inference, and mapped the detected poses to particle forces and camera movement. Everything runs client-side, so users can open the live demo, allow webcam access, and start interacting with the visualization.
In testing, the system reached real-time performance across desktop and mobile devices, with smooth particle control and minimal setup.