BSDPose is an open-source framework for badminton stroke recognition and quality assessment, integrating YOLOv8-Pose with CBAM attention and kinematic feature analysis. It supports six-class stroke classification and 0–100 quality scoring in a single multi-task pipeline.

📄 This repository accompanies the paper "Badminton Stroke Recognition and Quality Assessment Based on YOLOv8-Pose and Kinematic Feature Analysis" (湖南理工学院学报·自然科学版, 2026). The BSD dataset is derived from the VideoBadminton public benchmark.

🚀 Key Features

Feature	Description
🎯 CBAM-Enhanced YOLOv8-Pose	Channel + Spatial attention injected at PAN P3/P4/P5 outputs for better keypoint localization under motion blur
🏸 15-D Kinematic Features	Joint angles, torso inclination, wrist velocity, CoM displacement → statistical pooling over 16-frame windows → 62-D vector
🧠 Multi-Task Head	Shared encoder (62→128→64) + classification branch (6 strokes) + regression branch (0–100 quality), joint loss: CE + 0.5·MSE
📊 Full Experiment Pipeline	End-to-end: data preparation → pose training (baseline/SE/CBAM) → feature extraction → classifier training → evaluation

📦 Installation

Pip install (recommended)

git clone https://github.com/lechan775/BSDPose.git
cd BSDPose
pip install -r requirements.txt

Requirements: torch>=2.0, ultralytics>=8.0, opencv-python, pyyaml, numpy

🏃 Quick Start

1. Prepare BSD Dataset

# Download VideoBadminton (https://github.com/qilimk/VideoBadminton) first, then:
# Place .mp4 clips under data/pose/VideoBadminton_Dataset/
python scripts/prepare_bsd.py \
    --source-root data/pose \
    --output-root data/bsd \
    --window-size 16 \
    --seed 42

Output: data/bsd/{images,labels,train.json,val.json,test.json,bsd-pose.yaml}

2. Train Pose Estimation

# All three variants (baseline, SE, CBAM)
python train/train_pose.py --variant all --epochs 200 --batch 16

# Single variant
python train/train_pose.py --variant cbam --device 0

3. Train Classifier

# All classifier experiments
python train/train_classifier.py --variant all --epochs 200 --batch-size 64

# Our proposed method only
python train/train_classifier.py --variant cbam_kinematic_multitask

4. Evaluate

python eval/eval_pose.py        # Pose mAP on BSD test split
python eval/eval_classifier.py  # Accuracy / Precision / Recall / F1 / MAE
python eval/eval_ablation.py    # Full ablation results

📊 Benchmark Results

Pose Estimation (BSD Test Split)

Model	mAP@0.5:0.95 (%)
YOLOv8-Pose (baseline)	82.46
+ SE	82.19
+ CBAM (ours)	83.46

Stroke Classification (BSD Test Split)

Method	Accuracy (%)	Precision (%)	Recall (%)	F1-Score (%)
OpenPose + LSTM	45.79	44.00	45.79	43.49
YOLOv8-Pose + LSTM	46.15	44.00	46.15	43.29
YOLOv8-Pose + Kinematic + MLP	56.78	57.80	56.78	57.11
YOLOv8-Pose + CBAM + Kinematic + Multi-Task (ours)	57.78	58.80	57.78	58.11

Ablation Study

Pose Model	CBAM	Kinematic	Multi-Task	Accuracy (%)	Quality MAE
YOLOv8-Pose	✗	✗	✗	46.15	—
YOLOv8-Pose	✓	✗	✗	44.50	—
YOLOv8-Pose	✗	✓	✓	53.48	2.41
YOLOv8-Pose	✓	✓	✓	57.78	2.20

Key Findings

• CBAM alone does not improve classification when used with raw LSTM (44.50% vs 46.15%)
• CBAM + kinematic features synergy: combined use boosts accuracy by +4.30 pp (53.48% → 57.78%)
• Quality MAE reduction: multi-task head compresses MAE from 2.41 to 2.20

🧩 Project Structure

BSDPose/
├── models/                      # Attention modules & YOLOv8-Pose injection
│   ├── cbam.py                  #   ChannelAttention, SpatialAttention, CBAM, SEAttention
│   └── yolov8_cbam_pose.py      #   inject_pan_attention(), build_yolov8_pose_with_attention()
├── features/
│   └── kinematic.py             #   15 per-frame features → 62-D window aggregation
├── classifier/
│   ├── multi_task_head.py       #   MultiTaskHead, MultiTaskLoss, MLP/LSTM baselines
│   └── dataset.py               #   KinematicFeatureDataset, FeatureRecord loader
├── train/
│   ├── train_pose.py            #   Train baseline/SE/CBAM YOLOv8-Pose variants
│   └── train_classifier.py      #   Train LSTM/MLP/MultiTask classifiers
├── scripts/
│   └── prepare_bsd.py           #   VideoBadminton → BSD dataset builder
├── eval/
│   ├── eval_pose.py             #   Pose evaluation (mAP)
│   ├── eval_classifier.py       #   Classifier evaluation (accuracy, MAE)
│   └── eval_ablation.py         #   Ablation experiment runner
├── configs/
│   └── config.yaml              #   Hyperparameters & experiment matrix
└── paper/
    └── journal_draft.md         #   Manuscript (Markdown)

🔬 Methodology

VideoBadminton (.mp4)
        │
        ▼
  16-Frame Window Sampling
        │
        ▼
┌───────────────────────────────┐
│  YOLOv8-Pose Variants         │
│  ┌─────────┬────────┬───────┐ │
│  │Baseline │  + SE  │ +CBAM │ │
│  │82.46%   │ 82.19% │83.46% │ │
│  └─────────┴────────┴───────┘ │
└───────────────────────────────┘
        │
        ▼  COCO-17 Keypoints (x,y,conf)
┌───────────────────────────────────┐
│  Kinematic Feature Extraction      │
│  15 per-frame → 16-frame window    │
│  → 62-D vector (mean/std/max/min)  │
└───────────────────────────────────┘
        │
        ▼
┌───────────────────────────────────┐
│  Multi-Task Head                   │
│  Shared:  62 → 128 → 64           │
│  ┌──────────────┬────────────────┐ │
│  │ Classifier   │   Regressor    │ │
│  │ 64→32→6      │   64→32→1      │ │
│  │ (Softmax)    │   (Sigmoid)    │ │
│  └──────────────┴────────────────┘ │
│  L = CE + 0.5·MSE                  │
└───────────────────────────────────┘
        │
        ▼
  Output: 6-Class Stroke + 0–100 Quality Score

🏷️ Stroke Classes

#	English	中文
0	Forehand Clear	正手高远球
1	Backhand Clear	反手高远球
2	Forehand Smash	正手杀球
3	Backhand Smash	反手杀球
4	Forehand Net Lift	正手网前挑球
5	Backhand Net Lift	反手网前挑球

📖 Citation

@article{bsdpose2026,
  title     = {{基于YOLOv8-Pose与运动学特征分析的羽毛球击球动作识别与质量评估研究}},
  author    = {胡竞文 and 罗正 and 潘润杭 and 范飞豪},
  journal   = {湖南理工学院学报(自然科学版)},
  year      = {2026},
  note      = {Submitted}
}

📝 License

This project is licensed under the MIT License — free for academic and commercial use.

---

_{Built with ❤️ for Badminton AI Research}