Computer VisionResearch
Computer vision system · Prototype
ChessVision
Reached 90%+ move detection accuracy for real-time over-the-board chess digitization.
Overview
ChessVision captures physical chess games and converts them into a live digital record. The system links camera perception with chess-specific state tracking so real games can be logged in real time.
My contribution
YOLOv11 training pipeline, OpenCV board localization, and real-time game-state integration for TJHSST senior research.
Problem
Over-the-board chess is difficult to digitize reliably because board angles, lighting, and piece occlusion all destabilize frame-level detection.
Approach
- Trained a YOLOv11 model on Roboflow-annotated chess piece images for piece detection.
- Used OpenCV for live board localization before mapping detections into game-state updates.
- Built the system as part of TJHSST senior research with the goal of real-time recording rather than offline analysis.
Result
Reached 90%+ move detection accuracy across arbitrary board angles and turned live games into structured digital records.
Stack
PythonRoboflowOpenCVYOLOv11Tkinter