Getting reliable data effortless from volleyball recordings is a difficult task. This task can be simplified by the assistance of machine learning concepts such as Computer Vision or Convolutional Neural Networks. This thesis describes an applicable way that is feasible to implement for amateur teams that will not have the most advanced devices. The application developed while working on this topic is a proof-of-concept on how to extract impact points of serves from a recording or live stream that was filmed from a side view. The results are displayed on a top-down diagram of the court. The working solution of this ball tracking problem can easily be extended to support other use cases.
The goal of providing a usable framework for this kind of statistics gathering was reached, although the early results showed the need for different strategy for tracking the ball. The concept of using a Convolutional Neural Network to categorize the balls was dropped and instead a solution of detecting the ball with use of heavy preprocessing and a combination using image processing features and domain knowledge was developed.
This proved to be reliable in the given environment and resulted in few errors while detecting the ball.