BaddieVision: An Intelligent Badminton Video Analysis System Based on Computer Vision

BaddieVision is a complete badminton video analysis toolchain that integrates MediaPipe pose estimation, TrackNetV3 shuttlecock tracking, YOLO object detection, and LSTM temporal classification technologies to enable end-to-end automated analysis from raw video to tactical insights. The system can extract player poses, shuttlecock trajectories, shot types, and rally states, supporting various application scenarios.

• Original Author/Maintainer: MaxLinCode
• Source Platform: GitHub
• Original Title: BaddieVision: Computer vision and machine learning tools for badminton video analysis
• Original Link: https://github.com/MaxLinCode/BaddieVision
• Release Date: July 12, 2026

This project is specifically designed for badminton, aiming to automatically extract key information from match videos using computer vision technology and build a complete sports analysis pipeline.

1. Player Pose Extraction (MediaPipe)

Uses Google MediaPipe framework for real-time human pose estimation, detects key skeletal points, analyzes player positions, movement trajectories, and body postures, and outputs standardized court coordinate system data.

2. Shuttlecock Tracking (TrackNetV3 + YOLO)

Adopts a dual-track strategy: TrackNetV3 predicts shuttlecock position heatmaps to generate candidate trajectories; YOLO object detection assists in verification, solving the problem of small ball tracking.

3. Shot Type Classification

Identifies shot types such as clears, drops, smashes, etc., based on a 76-dimensional feature vector (66 pose coordinates +7 shuttlecock features +3 court anchor points).

4. Rally Detection (LSTM)

Performs frame-level rally state classification via LSTM network to determine if it is in an active match state, supporting highlight rally clipping and duration statistics.

Court Calibration

Provides a browser-based interactive tool where users click on court lines to calculate a homography matrix, mapping pixel coordinates to a standard metric coordinate system. It supports drag guidance, zoom/pan, magnifier assistance, and headless operation.

Data Processing Flow

1. Video Preprocessing: FFmpeg frame rate conversion and clipping
2. Feature Extraction: Parallel extraction of pose, ball trajectory, and court projection features
3. Data Alignment: Align features by timestamp to generate training samples
4. Model Training: Independently train shot classifier and in-play detector
5. Visualization Output: Overlay analysis results onto the original video

Reproducibility Design

• Seeded random numbers to ensure consistent results
• Training/validation sets grouped by video source to prevent data leakage
• Dynamic path parsing without relying on fixed directories

Cross-Platform Support

• CUDA (NVIDIA GPU acceleration)
• MPS (Apple Silicon backend)
• CPU general fallback

Data Management

Large resource files (videos, model weights, etc.) are ignored via Git, and a ZIP packaging solution is provided for easy collaboration.

BaddieVision can be applied to:

• Athlete training analysis: Quantify movement distance, reaction time, shot selection
• Tactical research: Identify opponents' usual patterns and weaknesses
• Automated referee assistance: Detect controversial balls and out-of-bounds judgments
• Event broadcast enhancement: Generate real-time tactical statistics and visual overlays
• Youth training: Data feedback helps beginners understand technical movements

BaddieVision is a typical case of technology stack integration in the sports technology field, proving that integrating multiple computer vision technologies is necessary to solve problems like small ball tracking in real-world scenarios. It is an excellent reference implementation for developers in the sports analysis field; its modular design allows independent reuse of components (e.g., the pose extraction module can be used for other sports, and the ball tracking approach can be migrated to scenarios like table tennis or tennis).