BallScope
Two cameras, one moving view of the match. AI-assisted football tracking, recording and analysis.
Overview
The idea
A fixed camera can cover the pitch, but the ball is often hard to follow at that scale. We built BallScope to explore how two cameras and object detection could produce a tighter, moving view while preserving the original footage.
- Built by
- Rafael Reverberi and Benjamin Flury
- Platforms
- Apple Silicon macOS · NVIDIA Jetson
- Status
- Research project
- License
- MIT
How it works
Inside the project
Live capture
Two parallel camera feeds run through YOLO detection. The system chooses a view and smooths the crop and zoom, while FastAPI serves browser previews and recording controls.
Offline analysis
Separate left and right uploads advance on a shared playback timeline. Detections are mapped into a stitched view so the export can follow play across the overlap.
Setup on the sideline
A camera workspace exposes live previews and saved presets. BallScope Remote adds an iPhone interface for supported capture and analysis controls.
Status and limitations
Setup depends on compatible cameras and the platform-specific video and inference runtimes.
