GoaliePose

Computer vision for ice hockey goalie movement analysis.

GoaliePose is a computer vision project focused on detecting and tracking a single goalie in fixed-camera video to generate a clean skeletal/keypoint representation for coaching and technique review. The long-term goal is to turn raw footage into usable movement insights for goalie coaches and self-training players.

Project Status

Current phase: Phase 0 — Product spec and benchmarks

This repository is currently in the product-definition and benchmark-building stage. The focus right now is to lock scope, define acceptance criteria, and create a small but meaningful benchmark set before model experimentation begins.

Problem

Goalie coaching often relies on manual video review, which is slow, subjective, and difficult to scale. This project aims to automate the first technical layer of review by identifying a goalie’s body position frame-by-frame and producing stable tracked keypoints that can later support deeper movement analysis.

A correct result means:

• the goalie is found reliably in the frame
• body keypoints are placed accurately enough to form a meaningful skeleton
• motion can be reviewed consistently across a clip

A poor result means:

• keypoints drift or disappear
• the skeleton becomes unusable for review
• the footage loses coaching value

Version 1 Scope

The initial version is intentionally narrow.

Included in v1

• Fixed camera input only
• Single goalie only
• Batch processing only
• Keypoint/body tracking for goalie posture and movement review
• Benchmark-driven development

Not included in v1

• Multi-camera support
• Live/real-time inference as a required feature
• Multiple players / full-rink understanding
• General hockey scene understanding
• Stick-state classification, unless the baseline proves stable enough

Phase 0 Goals

Phase 0 exists to prevent wasted model work by defining exactly what success looks like.

Deliverables

• One-page product spec
• Fixed ontology v1
• Acceptance criteria
• 20–30 benchmark clips selected manually
• 5–10 must-pass representative scenarios

Exit criteria

Phase 0 is complete when:

• the v1 scope is frozen
• ontology labels are fixed
• acceptance criteria are written
• benchmark clips are selected and organized
• representative must-pass scenarios are documented

Core Idea

Input:

• goalie practice footage from a fixed camera angle

Output:

• tracked body keypoints / skeletal representation of the goalie over time

Primary use case:

• post-session coaching review
• self-review by goalies
• movement and positioning analysis

Users

This project is being designed for:

• ice hockey goalie coaches
• self-teaching goalies
• players looking to review technique through video

Guiding Principles

• Narrow scope first. A small system that works is better than a broad system that fails.
• Benchmark before scale. Build around hard examples early.
• Stability over novelty. A reliable baseline matters more than flashy features.
• Coaching utility first. Output should be useful for review, not just technically impressive.

Proposed Repository Structure

goaliepose/
├── README.md
├── LICENSE
├── docs/
│   ├── product-spec.md
│   ├── ontology-v1.md
│   ├── acceptance-criteria.md
│   └── roadmap.md
├── benchmarks/
│   ├── clips/
│   ├── metadata/
│   └── scenarios/
├── data/
│   ├── raw/
│   ├── processed/
│   └── samples/
├── src/
│   ├── ingestion/
│   ├── detection/
│   ├── tracking/
│   ├── keypoints/
│   ├── evaluation/
│   └── visualization/
├── experiments/
├── notebooks/
└── tests/

Phase Roadmap

Phase 0 — Product spec and benchmarks

Define the problem, freeze v1 scope, and build the benchmark set.

Phase 1 — Baseline pipeline

Build the first working pipeline for:

• clip ingestion
• goalie detection / isolation
• keypoint estimation
• tracking
• visual overlay output

Later phases

Planned future work may include:

• stronger benchmark coverage
• robustness improvements
• better tracking stability
• richer analytics
• possible real-time or multi-camera extensions

Benchmark Philosophy

The benchmark set should reflect real coaching conditions, not just ideal footage.

Examples of scenarios to include:

• clean centered goalie in stable lighting
• lateral movement across the crease
• butterfly transitions
• partial occlusion
• darker rink lighting
• varying goalie gear / jersey contrast
• camera distance changes
• fast recovery movements

The project should be judged against must-pass scenarios, not just average-case success.

Acceptance Direction

The exact numbers will be finalized in Phase 0, but the system should eventually be evaluated on qualities like:

• keypoint stability
• tracking continuity
• robustness to lighting and motion
• coaching usefulness of output
• failure rate on benchmark scenarios

Current Priorities

1. Write the one-page product spec
2. Finalize ontology v1
3. Define acceptance criteria
4. Curate the benchmark clip set
5. Identify the must-pass scenarios
6. Begin baseline pipeline design

License

This project is licensed under the Apache License 2.0. See the LICENSE file for details.

Notes

This repository is under active development. During the early phases, the priority is correctness of scope and evaluation setup rather than feature breadth.