GitHub - wish44165/A-New-Perspective-for-Shuttlecock-Hitting-Event-Detection: IJCAI 2023 (Macao, S.A.R)

A New Perspective for Shuttlecock Hitting Event Detection

AICUP2023: Teaching Computer to Watch Badminton Matches - Taiwan's first competition combining AI and sports (TEAM_3009)

CodaLab2023: Track 1 Automatic Annotation of Technical Data for Badminton Match Videos (Team name: yc)

International Workshop on Intelligent Technologies for Precision Sports Science (IT4PSS) in Conjunction with IJCAI 2023, the 32nd International Joint Conference on Artificial Intelligence

CoachAI Badminton Challenge 2023 (Team name: yc)

$\large{\textbf{Abstract}}$

This article introduces a novel approach to shuttlecock-hitting event detection. Instead of depending on generic methods, we capture the hitting action of players by reasoning over a sequence of images. To learn the features of hitting events in a video clip, we specifically utilize a deep learning model known as SwingNet. This model is designed to capture the relevant characteristics and patterns associated with the act of hitting in badminton. By training SwingNet on the provided video clips, we aim to enable the model to accurately recognize and identify the instances of hitting events based on their distinctive features. Furthermore, we apply the specific video processing technique to extract the prior features from the video, significantly reducing the model's learning difficulty. The proposed method not only provides an intuitive and user-friendly approach but also presents a fresh perspective on the task of detecting badminton-hitting events. The datasets and trained models can be found here.

1. Environmental Setup

Hardware Information

CPU: Intel® Core™ i7-11700F
GPU: GeForce GTX 1660 SUPER™ VENTUS XS OC (6G)

Create Conda Environments

TrackNetv2

$ conda create -n tracknetv2 python=3.9 -y

SwingNet

$ conda create -n golfdb python=3.8 -y

ViT

$ conda create -n ViT_j python==3.9 -y

YOLOv5

$ conda create -n yolov5 python=3.7 -y

YOLOv8

$ conda create -n yolov8 python=3.7 -y

Install Required Packages

TrackNetv2

$ conda activate tracknetv2
$ git clone https://nol.cs.nctu.edu.tw:234/lukelin/TrackNetV2_pytorch.git
$ sudo apt-get install git
$ sudo apt-get install python3-pip
$ pip3 install pandas
$ pip3 install opencv-python
$ pip3 install matplotlib
$ pip3 install -U scikit-learn
$ pip3 install torch
$ pip3 install torchvision

SwingNet

$ conda activate golfdb
$ git clone https://github.com/wmcnally/golfdb.git
$ pip3 install opencv-python
$ pip3 install scipy
$ pip3 install pandas
$ pip3 install torch
$ pip3 install torchvision
$ pip3 install torchaudio

ViT

$ conda activate ViT_j
$ git clone https://github.com/jeonsworld/ViT-pytorch.git
$ cd ViT-pytorch/
$ pip3 install -r requirements.txt
$ mkdir checkpoint/
$ cd checkpoint/
$ wget https://storage.googleapis.com/vit_models/imagenet21k+imagenet2012/ViT-B_16.npz
$ git clone https://github.com/NVIDIA/apex    # A PyTorch Extension
$ cd apex/
$ python3 setup.py install

YOLOv5

$ conda activate yolov5
$ git clone https://github.com/ultralytics/yolov5.git
$ cd yolov5/
$ pip install -r requirements.txt

YOLOv8

$ conda activate yolov8
$ git clone https://github.com/ultralytics/ultralytics.git
$ cd ultralytics/
$ pip install -r requirements.txt

for Conda users

$ cd crc/colab/
$ conda env create -f environment.yml
$ conda activate shuttlecock

2. Inference Details

Datasets

The datasets and trained models can be found here.

Stage 1 dataset

Stage 2 dataset

Folder Structure on Local Machine

Create the following folder structure on the local machine

Badminton/
├── data/
    └── part1/
        └── val/
└── src/
    ├── TrackNetV2_pytorch/
        ├── 10-10Gray/
            ├── denoise10_custom.py
            └── predict10.py
        ├── HitFrame.py
        ├── LandingX.py
        └── event_detection_custom.py
    ├── ultralytics/
        ├── demo.py
        └── submit.py
    ├── ViT-pytorch_Backhand/
        └── submit.py
    ├── ViT-pytorch_BallHeight/
        └── submit.py
    ├── ViT-pytorch_BallType/
        └── submit.py
    ├── ViT-pytorch_Hitter/
        └── submit.py
    ├── ViT-pytorch_RoundHead/
        └── submit.py
    ├── ViT-pytorch_Winner/
        └── submit.py
    ├── postprocess/
        ├── get_hitframe_yolo.py
        └── get_hitframe.py
    ├── preprocess/
        └── rt_conversion_datasets.py
    └── yolov5/
        ├── LandingY_Hitter_Defender_Location.py
        ├── demo.py
        └── detect.py

VideoName, ShotSeq, HitFrame

put Badminton/data/part2/test/00170/ .. /00399/ into Badminton/data/part1/val/

→ Badminton/data/part1/val/00001/ .. /00399/    # 1280x720
# CodaLab
→ Badminton/data/CodaLab/testdata_track1/00170/ .. /00399/    # 1280x720

convert val/+test/ to val_test_xgg/

$ conda activate golfdb
$ cd Badminton/src/preprocess/
$ mkdir val_test_xgg
$ python3 rt_conversion_datasets.py
→ Badminton/src/preprocess/val_test_xgg/    # 1280x720
# CodaLab
→ Badminton/src/preprocess/CodaLab/testdata_track1/    # 1280x720

upload val_test_xgg/ to google drive Teaching_Computer_to_Watch_Badminton_Matches_Taiwan_first_competition_combining_AI_and_sports/datasets/part1/

→ Teaching_Computer_to_Watch_Badminton_Matches_Taiwan_first_competition_combining_AI_and_sports/datasets/part1/val_test_xgg/
→ execute golfdb_xgg_inference_best.ipynb
→ src/Notebook/golfdb/golfdb_G3_fold5_iter3000_val_test_X.csv    # 0.0426
# CodaLab
→ src/Notebook/golfdb/CodaLab_testdata_track1.csv

Hitter

put golfdb_G3_fold5_iter3000_val_test_X.csv into Badminton/src/postprocess/

→ Badminton/src/postprocess/golfdb_G3_fold5_iter3000_val_test_X.csv
# CodaLab
→ Badminton/src/postprocess/CodaLab/CodaLab_testdata_track1.csv

extract hitframe from csv file

$ cd Badminton/src/postprocess/
$ mkdir HitFrame
$ mkdir HitFrame/1
$ python3 get_hitframe.py
>> len(vns), len(hits), len(os.listdir(savePath)) = 4007, 4007, 4007
→ Badminton/src/postprocess/HitFrame/1/    # 720x720, 4007; # CodaLab: 720x720, 2408

execute hitter inference

$ conda activate ViT_j
$ cd Badminton/src/ViT-pytorch_Hitter/
$ python3 submit.py --model_type ["ViT-B_16","ViT-B_16","ViT-B_16","ViT-B_16","ViT-B_16"] --checkpoint ["output/fold1_Hitter_ViT-B_16_checkpoint.bin","output/fold2_Hitter_ViT-B_16_checkpoint.bin","output/fold3_Hitter_ViT-B_16_checkpoint.bin","output/fold4_Hitter_ViT-B_16_checkpoint.bin","output/fold5_Hitter_ViT-B_16_checkpoint.bin"] --img_size [480,480,480,480,480]
→ Badminton/src/ViT-pytorch_Hitter/golfdb_G3_fold5_iter3000_val_test_hitter_vote.csv    # 0.0494
→ Badminton/src/ViT-pytorch_Hitter/golfdb_G3_fold5_iter3000_val_test_hitter_mean.csv    # 0.0494
# CodaLab
→ Badminton/src/ViT-pytorch_Hitter/CodaLab_testdata_track1_hitter_vote.csv
→ Badminton/src/ViT-pytorch_Hitter/CodaLab_testdata_track1_hitter_mean.csv

RoundHead

execute roundhead inference

$ cd Badminton/src/ViT-pytorch_RoundHead/
$ python3 submit.py --model_type ["ViT-B_16","ViT-B_16","ViT-B_16","ViT-B_16","ViT-B_16"] --checkpoint ["output/fold1_RoundHead_ViT-B_16_checkpoint.bin","output/fold2_RoundHead_ViT-B_16_checkpoint.bin","output/fold3_RoundHead_ViT-B_16_checkpoint.bin","output/fold4_RoundHead_ViT-B_16_checkpoint.bin","output/fold5_RoundHead_ViT-B_16_checkpoint.bin"] --img_size [480,480,480,480,480]
→ Badminton/src/ViT-pytorch_Hitter/golfdb_G3_fold5_iter3000_val_test_hitter_vote_roundhead_vote.csv    # 0.0527
→ Badminton/src/ViT-pytorch_Hittergolfdb_G3_fold5_iter3000_val_test_hitter_mean_roundhead_mean.csv    # 0.0527
# CodaLab
→ Badminton/src/ViT-pytorch_RoundHead/CodaLab_testdata_track1_hitter_vote_roundhead_vote.csv
→ Badminton/src/ViT-pytorch_RoundHead/CodaLab_testdata_track1_hitter_mean_roundhead_mean.csv

Backhand

execute backhand inference

$ cd Badminton/src/ViT-pytorch_Backhand/
$ python3 submit.py --model_type ["ViT-B_16","ViT-B_16","ViT-B_16","ViT-B_16","ViT-B_16"] --checkpoint ["output/fold1_Backhand_ViT-B_16_checkpoint.bin","output/fold2_Backhand_ViT-B_16_checkpoint.bin","output/fold3_Backhand_ViT-B_16_checkpoint.bin","output/fold4_Backhand_ViT-B_16_checkpoint.bin","output/fold5_Backhand_ViT-B_16_checkpoint.bin"] --img_size [480,480,480,480,480]
# CodaLab
→ Badminton/src/ViT-pytorch_Backhand/CodaLab_testdata_track1_hitter_vote_roundhead_vote_backhand_vote.csv
→ Badminton/src/ViT-pytorch_Backhand/CodaLab_testdata_track1_hitter_mean_roundhead_mean_backhand_mean.csv

BallHeight

execute ballheight inference

$ cd Badminton/src/ViT-pytorch_BallHeight/
$ python3 submit.py --model_type ["ViT-B_16","ViT-B_16","ViT-B_16","ViT-B_16","ViT-B_16"] --checkpoint ["output/fold1_BallHeight_ViT-B_16_checkpoint.bin","output/fold2_BallHeight_ViT-B_16_checkpoint.bin","output/fold3_BallHeight_ViT-B_16_checkpoint.bin","output/fold4_BallHeight_ViT-B_16_checkpoint.bin","output/fold5_BallHeight_ViT-B_16_checkpoint.bin"] --img_size [480,480,480,480,480]
# CodaLab
→ Badminton/src/ViT-pytorch_BallHeight/CodaLab_testdata_track1_hitter_vote_roundhead_vote_backhand_vote_ballheight_vote.csv
→ Badminton/src/ViT-pytorch_BallHeight/CodaLab_testdata_track1_hitter_mean_roundhead_mean_backhand_mean_ballheight_mean.csv

LandingX

get trajectory

$ conda activate tracknetv2
$ cd Badminton/src/TrackNetV2_pytorch/10-10Gray/
$ mkdir output
$ python3 predict10_custom.py
$ mkdir denoise
$ python3 denoise10_custom.py

execute landingx inference

$ cd Badminton/src/TrackNetV2_pytorch/10-10Gray/
$ (mkdir event
$ cd Badminton/src/TrackNetV2_pytorch/
$ python3 event_detection_custom.py
$ python3 HitFrame.py)
# CodaLab
→ Badminton/src/TrackNetV2_pytorch/CodaLab_tracknetv2_pytorch_10-10Gray_denoise_eventDetection_X.csv
$ python3 LandingX.py
# CodaLab
→ Badminton/src/TrackNetV2_pytorch/CodaLab_testdata_track1_hitter_vote_roundhead_vote_backhand_vote_ballheight_vote_LXY.csv
→ Badminton/src/TrackNetV2_pytorch/CodaLab_testdata_track1_hitter_mean_roundhead_mean_backhand_mean_ballheight_mean_LXY.csv

LandingY, HitterLocationX, HitterLocationY, DefenderLocationX, DefenderLocationY

extract hitframe for yolo from csv

$ cd Badminton/src/postprocess/
$ mkdir HitFrame_yolo
$ python3 get_hitframe_yolo.py
→ Badminton/src/postprocess/HitFrame_yolo/    # 1280x720, 4007; CodaLab: 1280x720, 2408

execute yolov5 inference

$ conda activate yolov5
$ cd Badminton/src/yolov5/
$ python3 detect.py --weights runs/train/exp/weights/best.pt --source /home/yuhsi/Badminton/src/postprocess/HitFrame_yolo/ --conf-thres 0.3 --iou-thres 0.3 --save-txt --imgsz 2880 --agnostic-nms --augment
→ Badminton/src/yolov5/runs/detect/exp/    # 4007
# CodaLab
$ python3 detect.py --weights runs/train/exp/weights/best.pt --source /home/yuhsi/Badminton/src/postprocess/HitFrame_yolo/ --conf-thres 0.3 --iou-thres 0.3 --save-txt --imgsz 2880 --agnostic-nms --augment
→ Badminton/src/yolov5/runs/detect/exp2/    # 2408
## video demo
$ python3 detect.py --weights runs/train/exp/weights/best.pt --source /home/yuhsi/Badminton/data/CodaLab/testdata_track1/00171/00171.mp4 --conf-thres 0.3 --iou-thres 0.3 --save-txt --imgsz 2880 --agnostic-nms --augment
$ python3 demo.py

execute landingy inference

$ mkdir runs/detect/exp_draw
$ mkdir runs/detect/exp_draw/case1
$ python3 LandingY_Hitter_Defender_Location.py

BallType

execute balltype inference

$ conda activate ViT_j
$ cd Badminton/src/ViT-pytorch_BallType/
$ python3 submit.py --model_type ["ViT-B_16","ViT-B_16","ViT-B_16","ViT-B_16","ViT-B_16"] --checkpoint ["output/fold1_BallType_ViT-B_16_checkpoint.bin","output/fold2_BallType_ViT-B_16_checkpoint.bin","output/fold3_BallType_ViT-B_16_checkpoint.bin","output/fold4_BallType_ViT-B_16_checkpoint.bin","output/fold5_BallType_ViT-B_16_checkpoint.bin"] --img_size [480,480,480,480,480]
# CodaLab
→ Badminton/src/ViT-pytorch_BallType/CodaLab_testdata_track1_hitter_mean_roundhead_mean_backhand_mean_ballheight_mean_LX_LY_case1_HD_balltype_vote.csv
→ Badminton/src/ViT-pytorch_BallType/CodaLab_testdata_track1_hitter_vote_roundhead_vote_backhand_vote_ballheight_vote_LX_LY_case1_HD_balltype_mean.csv

Winner

execute winner inference

$ cd Badminton/src/Vit-pytorch_Winner/
$ python3 submit.py --model_type ["ViT-B_16","ViT-B_16","ViT-B_16","ViT-B_16","ViT-B_16"] --checkpoint ["output/fold1_Winner_ViT-B_16_checkpoint.bin","output/fold2_Winner_ViT-B_16_checkpoint.bin","output/fold3_Winner_ViT-B_16_checkpoint.bin","output/fold4_Winner_ViT-B_16_checkpoint.bin","output/fold5_Winner_ViT-B_16_checkpoint.bin"] --img_size [480,480,480,480,480]
# CodaLab
→ Badminton/src/ViT-pytorch_Winner/CodaLab_testdata_track1_hitter_mean_roundhead_mean_backhand_mean_ballheight_mean_LX_LY_case1_HD_balltype_vote_winner_mean_case1.csv

HitterLocationX, HitterLocationY, DefenderLocationX, DefenderLocationY (Updated)

use yolov8x-pose-p6.pt model to execute pose estimation

$ cd Badminton/src/ultralytics/
$ mkdir pose_estimation
$ python3 submit.py
→ Badminton/src/ViT-pytorch_Winner/CodaLab_testdata_track1_hitter_mean_roundhead_mean_backhand_mean_ballheight_mean_LX_LY_case1_HD_balltype_vote_winner_mean_case1_v8pose.csv
## video demo
$ python3 demo.py

3. Demonstration

3.1. Optical Flow Calculation

3.2. SwingNet (MobileNetV2 + bidirectional LSTM)

3.3. YOLOv5 & TrackNetV2 & YOLOv8-pose

4. Leaderboard Scores

4.1. AICUP2023

Leaderboards	Filename	Upload time	Evaluation result	Ranking
Public	golfdb_G3_fold5_...csv	2023-05-15 22:21:17	0.0727	11/30
Private	golfdb_G3_fold5_...csv	2023-05-15 22:21:17	0.0622	11/30

4.2. CodaLab2023

Leaderboards	Filename	Upload time	Evaluation result	Ranking
Final phase	CodaLab_testdata_track1_...csv	2023-06-17 16:03	0.3483	2/2

5. GitHub Acknowledgement

6. References

Citation

If you find this project helpful for your research or applications, we would appreciate it if you could give it a star and cite the paper.

@misc{chen2023new,
      title={A New Perspective for Shuttlecock Hitting Event Detection}, 
      author={Yu-Hsi Chen},
      year={2023},
      eprint={2306.10293},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}

@misc{chen_2025_14677727,
  author       = {Chen, Yu-Hsi},
  title        = {A New Perspective for Shuttlecock Hitting Event
                   Detection
                  },
  month        = jan,
  year         = 2025,
  publisher    = {Zenodo},
  version      = {1.0.0},
  doi          = {10.5281/zenodo.14677727},
  url          = {https://doi.org/10.5281/zenodo.14677727},
}

Name		Name	Last commit message	Last commit date
Latest commit History 151 Commits
assets		assets
src		src
submit		submit
README.md		README.md

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Repository files navigation

A New Perspective for Shuttlecock Hitting Event Detection

AICUP2023: Teaching Computer to Watch Badminton Matches - Taiwan's first competition combining AI and sports (TEAM_3009)

CodaLab2023: Track 1 Automatic Annotation of Technical Data for Badminton Match Videos (Team name: yc)

International Workshop on Intelligent Technologies for Precision Sports Science (IT4PSS) in Conjunction with IJCAI 2023, the 32nd International Joint Conference on Artificial Intelligence

CoachAI Badminton Challenge 2023 (Team name: yc)

1. Environmental Setup

TrackNetv2

SwingNet

ViT

YOLOv5

YOLOv8

TrackNetv2

SwingNet

ViT

YOLOv5

YOLOv8

for Conda users

2. Inference Details

Stage 1 dataset

Stage 2 dataset

3. Demonstration

3.1. Optical Flow Calculation

3.2. SwingNet (MobileNetV2 + bidirectional LSTM)

3.3. YOLOv5 & TrackNetV2 & YOLOv8-pose

4. Leaderboard Scores

4.1. AICUP2023

4.2. CodaLab2023

5. GitHub Acknowledgement

6. References

Citation

About

Releases

Packages

Languages

wish44165/A-New-Perspective-for-Shuttlecock-Hitting-Event-Detection

Folders and files

Latest commit

History

Repository files navigation

A New Perspective for Shuttlecock Hitting Event Detection

AICUP2023: Teaching Computer to Watch Badminton Matches - Taiwan's first competition combining AI and sports (TEAM_3009)

CodaLab2023: Track 1 Automatic Annotation of Technical Data for Badminton Match Videos (Team name: yc)

International Workshop on Intelligent Technologies for Precision Sports Science (IT4PSS) in Conjunction with IJCAI 2023, the 32nd International Joint Conference on Artificial Intelligence

CoachAI Badminton Challenge 2023 (Team name: yc)

1. Environmental Setup

TrackNetv2

SwingNet

ViT

YOLOv5

YOLOv8

TrackNetv2

SwingNet

ViT

YOLOv5

YOLOv8

for Conda users

2. Inference Details

Stage 1 dataset

Stage 2 dataset

3. Demonstration

3.1. Optical Flow Calculation

3.2. SwingNet (MobileNetV2 + bidirectional LSTM)

3.3. YOLOv5 & TrackNetV2 & YOLOv8-pose

4. Leaderboard Scores

4.1. AICUP2023

4.2. CodaLab2023

5. GitHub Acknowledgement

6. References

Citation

About

Topics

Resources

Stars

Watchers

Forks

Releases

Packages 0

Languages

Packages