Event Detection in Commertial Rugby Footage

A dissertation project by Mark Newell to detect events such as rucks, mauls and scrums from commertial rugby footage using computer vision techniques. Uses parts from:

1. Semseg - https://github.com/hszhao/semseg
2. AlphaPose - https://github.com/MVIG-SJTU/AlphaPose
3. Yolov3 - https://github.com/pjreddie/darknet

Prerequisits

1. Python 3.6.4
2. CUDA 10

Setup

The following setup is for a local environment. If you do however wish to run this in google colab (where most of development and testing took place) the .ipynb file has been provided.

Setup Environment

git clone https://github.com/markcNewell/RugbyEventDetection.git
cd RugbyEventDetection
pip install -r requirements.txt
mkdir "./out"
mkdir "./trained"
mkdir "./trained/pspnet"
mkdir "./trained/classification"

Install AlphaPose

git clone https://github.com/markcNewell/AlphaPose.git
mv AlphaPose alphapo
cd alphapo/
export PATH=/usr/local/cuda/bin/:$PATH
export LD_LIBRARY_PATH=/usr/local/cuda/lib64/:$LD_LIBRARY_PATH
sudo apt-get install libyaml-dev
python setup.py build develop --user

Install Darknet

git clone https://github.com/markcNewell/darknet.git
cd darknet
make

Once the packages have all built, the pretrained models need to be downloaded from the model storage:

• Put the train_epoch_100.pth file in RugbyEventDetection/trained/pspnet/
• Unzip the dataset dataset.zip to RugbyEventDetection
• Put the yolov3-spp.weights file in RugbyEventDetection/alphapo/detector/yolo/data/
• Put the fast_421_res152_256x192.pth in RugbyEventDetection/alphapo/pretrained_models/
• Put the model.cfg file in RugbyEventDetection/darknet/model/
• Put the model.weights file in RugbyEventDetection/darknet/model/

Getting Started

A runner.sh script has been provided to showcase the running of the algorithm but it can be done by using the following commands:

• python training.py --config config/training.yaml
• python main.py --config config/config-seg.yaml

As the processing takes a number of minutes for a short video an example has been provided () the results.json and video file should be downloaded. A script has been provided to combine the two. postprocessor.py should be run with the flags --json, --config and --video pointing to the respecive file locations.