Sports Vision
An automated pipeline that transforms raw sports broadcasts into fully tracked, annotated analysis sequences.
Raw broadcast video is unstructured and time-consuming to analyze manually. Coaches and analysts spend hours scrubbing through footage, tagging players, and estimating positions, slowing down the extraction of tactical insights.
A pipeline that ingests standard broadcast video and automatically detects every player and the ball. It tracks individuals across frames, classifies them into teams, and renders annotated output with trajectories and stats overlays.
Note: The health issues highlighted in the system HUD during this clip are due to incorrect player velocities—an artifact of the ongoing work in the Geometry & Motion projection step.
Methodology
Detection & Tracking
The pipeline detects every person and the ball in each frame using a modern object detection model. A tracking algorithm then assigns a persistent ID to each player as they move through the scene, maintaining identity even during occlusions or when players leave and re-enter the frame.
Team Classification
Players are classified into teams based on their visual appearance — primarily kit colours and design. The system builds a reference profile for each team and assigns stable labels to tracked players, handling lighting changes and occlusions without flickering between assignments.
Metrics & Analysis WIP
The next phase projects player positions from the video frame onto a real-world pitch to compute metrics like speed, distance covered, and positioning heatmaps. This step is currently under development.