← All projects

Sports Analytics · Indoor Hockey

Hockey Vision — tactical map

Automated player detection and tracking for indoor hockey — five classes tracked from a single fisheye stream and projected onto a live 2D tactical map with real-time team heatmaps.

ROLECV Engineer TYPEPersonal R&D STATUSV2 released
Hockey Vision: per-team distribution heatmaps, live match view, and 2D tactical map

The challenge

Five classes, one rink

Two teams, both goalkeepers, and the referee — visually similar, constantly crossing paths.

Fisheye geometry

A single wide lens warps the rink — raw positions are unusable without undistortion.

Pixels aren't tactics

Detections only become insight once they land on a 2D map a coach can actually read.

The approach

01

Dataset construction

Built the dataset in Roboflow with polygon-level annotations across all five classes.

02

YOLOv11 detection & tracking

Ultralytics YOLOv11 detecting and tracking teams, goalkeepers, and referee.

03

Fisheye undistortion

cv2.fisheye correction so the stream's geometry matches the real rink.

04

2D tactical mapping

Perspective projection, upgraded in V2 with field keypoints and a real-time heatmap.

Outcomes

5 classes

Detected & tracked

Both teams, both goalkeepers, and the referee.

2D map

Keypoint positioning

V2 mapping anchored to field keypoints for dynamic accuracy.

Real-time

Team heatmaps

Live positional heat for both teams on the pitch.

Stack YOLOv11 Ultralytics Roboflow Polygon annotation cv2.fisheye Perspective projection Field keypoints Real-time heatmap