StanlyHardy / open-tennis

Tennis match analysis via computer vision techniques.

OpenTennis

OpenTennis is a framework under active development to analyze Tennis matches. Currently, it supports cour edge extraction, and player information extraction via scoreboard analysis.

System Architecture • Features • Demo • Installation • Inference • Configurations • Roadmap

System Architecture

Features

• Court Edge Detection
• Recognize the Player Names.
• Determine the scores.
• Find the current serving player.
• Evaluate the average correct match.

Demo

Installation

Requirements

• Linux
• CUDA>= 10.0
• Python >= 3.6

Steps

1. Create a virtual conda environment and activate it.

conda create -n scorewatch python=3.9 -y
conda activate scorewatch

2. Install Pytorch

conda install pytorch torchvision torchaudio cudatoolkit=10.2 -c pytorch

4. Install TensorRT

pip install -U nvidia-tensorrt --index-url https://pypi.ngc.nvidia.com

5. Clone the repository

git clone https://github.com/StanlyHardy/score_watch # clone
cd score_watch

6. Install other requirements

pip install -r requirements.txt # install

7. The provided .engine file is platform specifc. So, export detector.pt within assets/models to TensorRT engine using the official exporter .

Inference

Inference could run either on Video or Image streams. The configuration could be changed via assets/config/app_config.yaml. If the evaluation is set to true, the inference occurs in validatation dataset and performs evaluation to determine the Average scores for correct Player names, Scores and Serving Player. Please change the input paths of video or images.

python app.py 

Configurations

App configuration(click to expand)

Section	Feature	Description
Paths	video_path	Path of the video on which the evaluation needs to be done.
	img_path	Directory containing the test images. Ground truth needs to be available for evaluation with image set.
	players_path	Path containing player informations
	groundtruth_path	Ground truth data which is in json format that has got the player information.
	output_video_path	The path to save the video if the output needs to be saved and visualized later.
	logs_path	Path where the output log will be saved.
Streamer	should_draw'	Draws over the frames for visualization , if enabled.
	view_imshow	The output visualization shall be turned on/off with this parameter.
	save_stream	Turning on this field enables the video output to be saved in the path defined in output_video_path
	debug	Displays debug logs if enabled
	evaluation	Turn on if the evaluation has to be done over the image set. Both image set and the annotations are required in this case.
Models	score_det_model'	Path of the score detector model.
	detector_config	Path of the config file for the score detector.
	text_rec_model	CRNN Model path responsible for Player information recognition.
	text_rec_config	Path to the configuration for the CRNN model
	ocr_engine	Choose between CRNN or PyTesseract.

Detector Configuration(click to expand)

Section	Feature	Description
YOLOv5	execution_env	ONNX Runtime provides support for CUDA, CPU and TensorRT. By default, CUDA is chosen. ONNX Runtime falls back to cpu if CUDA is unavailable.
	conf_thresh	Detection confidence
	iou_thres	IOU threshold to gauge the overlap.
	warm_up	Number of samples to be used during the warm up phase.
	class_labels	Class labels

Roadmap

• Train CRNN with wide set of Data from ATP/Wimbledon matches.
• Implement Ball Tracking, Trajectory Analysis
• Player tracking.
• Predict the style and the outcome of shot
• Player activity analysis

Acknowledgement

• ONNX Runtime 
• YOLOv5 
• TesserOCR 
• CRNN