Hockey-Vision CV Models

This repository contains five YOLO models trained on broadcast NHL video. These models are designed to be small and efficient (~2.5M parameters, ~5MB each), capable of running in real-time on most devices.

Together, they identify players, goalies, and referees, track the puck, read jersey numbers, and detect rink markings for top-down homography mapping.

Model	File	What it finds
Player	new_player_model.pt	Skaters, goalies, and referees
Puck	new_nano_puck.pt	The puck
Number	new_nums.pt	Jersey digits for player identification
Rink	new_rink_model.pt	Lines, circles, goal frame, and creases
Dots	new_dots.pt	Faceoff dots

---

Model Overviews & Examples

The examples below show each model running on three 10-second clips from a game between the Montreal Canadiens and San Jose Sharks. Click any image to view the full example video.

Player Model

new_player_model.pt

Finds every person on the ice and identifies their team. Classes include team_a_player, team_b_player, goalie_a, goalie_b, and referee. Teams are identified by jersey colors.

Clip 1	Clip 2	Clip 3

Puck Model

new_nano_puck.pt

A specialized detector for the puck. Because the puck is small and often obscured, this model is typically run with a higher confidence threshold (~0.6).

Clip 1	Clip 2	Clip 3

Number Model

new_nums.pt

A digit detector (0-9) that runs on player crops. It identifies jersey numbers through a multi-step process:

1. Digit Detection: Finds single digits on upscaled player crops.
2. Assembly: Groups neighboring digits into one- or two-digit numbers.
3. Temporal Voting: Aggregates readings across frames to settle on a confirmed number for each tracked player.

Clip 1	Clip 2	Clip 3

Rink Model

new_rink_model.pt

Detects painted rink markings (Blue Lines, Center Line, Circles, Goal Posts, and Creases). These detections are used to calculate the homography between the camera view and a top-down rink coordinate system.

Clip 1	Clip 2	Clip 3

Dots Model

new_dots.pt

Finds the faceoff dots. Since dots are precise points on the ice, they provide high-confidence anchors for the rink homography mapping.

Clip 1	Clip 2	Clip 3

---

Technical Details

These models are part of the larger Hockey-Vision system, which handles tracking (ByteTrack), homography mapping, and roster integration.

• Architecture: YOLOv8n (optimized as YOLO26n)
• Parameter Count: ~2.5M per model
• Inference Size: ~5MB per model

These models are used in my GitHub project: https://github.com/AlexRHaigh/Hockey-Vision