BowlingDL: A Deep Learning-Based Bowling Players Pose Estimation and Classification

Abstract

Human Pose Estimation (HPE) is one of the trending areas of research among artificial intelligent research. It has gained a lot of attention due to its versatile potential applications in various domains including transportation, healthcare, gaming, augmented reality, and sports. HPE can be used to build sports analytics, personalized training, and selflearning systems which allow players, athletes, and trainers to improve the training quality by evaluating various human poses detected from images or videos. As far as we know none of the exciting works considered developing a pose estimation and classification framework for bowling players. Therefore, in this paper, we proposed a deep-learning approach for bowling players’ pose estimation and classification. It uses our proposed Bowling Deep-Learning (BowlingDL) model along with the MoveNet model for bowling players’ pose estimation and classification. The MoveNet model detects various key points in human pose and the BowlingDL model classifies the detected bowling player’s poses into five different classes. For model training and evaluation, we collected and labelled our own dataset as no dataset was found for bowling posture. Our proposed model achieved 80% accuracy for the training dataset and 83% accuracy for the testing dataset. In addition, a smart mobile application for bowling players was developed where an edge-friendly version of BowlingDL–generated using TensorFlow Lite–was deployed.