Abstract
In this study, a custom neural network based on the DeepPose architecture is used to estimate human pose on sets of images containing images of people playing sports. Human pose estimation is studied, and deep neural network methods are introduced as a way of solving the human pose estimation problem as a regression problem. Through training on the Leeds Sports Pose Dataset and evaluating the custom neural network with the two common pose metrics, it was found that the initial deep neural network-based architecture used as the feature extractor is prone to overfitting due to the large number of model parameters. This led to the addition of dropout layers and adjusting the training length, which resulted in slight improvements in the estimation accuracy.