Soccer_player_tracking using Tensorflow 2.0

YOLOv3 Object Detection Model trained on Custom dataset

An implementation to track Soccer players on any random video. The model also detects background such as crowd and hoardings.

Project Requirements

• Python 3.7
• TensorFlow 2.0
• OpenCV
• Numpy
• Matplotlib
• Lxml
• tqdm
• Cudnn
• Cuda toolkit

Usage

1. Creating an environment for the repository using conda(Installation)

Tensorflow CPU

conda env create -f conda-cpu.yml // To create an environment where tensorflow-gpu is not supported

conda activate soccer-cpu

Tensorflow GPU

conda env create -f conda-gpu.yml // Creating an environment for tensorflow-gpu

conda activate soccer-gpu

2. Downloading weights

• Download the trained weights from here
• Move the downloaded weights and files to checkpoints/

Detections

for image detection

python detect.py --image ./data/socgirl.jpg // for detections in an image

for video

python detect_video.py --video ./data/goals.mp4 // for detections in a video file

Output

File Details

convert.py // Convert the weights of YOLOv3 to .tf format

train.py // to train your own model using custom dataset

utils.py // Draws outputs onto the image/frame using information received from model

models.py // All the model functions are in here.

coco.names // Contains the class names of the COCO dataset

soccerv2.names // Contains the class names of our custom trained model

References

YOLOv3 implemented in TensorFlow 2.0 by Zihao Zhang