A Case Study on Improving Ball Tracking Performance through Error Analysis Based on Data Class Refinement

This study addresses performance issues in a legacy golf ball tracker deployed in industrial applications, where recognition failures were long suspected to be caused by environmental factors such as ball color, obstacles, and lighting changes. This study presents a progressive class refinement method to systematically identify these error sources and leverages the findings to build a reliable, deep learning-based tracker. To investigate failures on a data-driven basis, this study employed this refinement method to define various environmental classes and analyze the legacy tracker’s errors within each class. The analysis confirmed that the tracker’s vulnerability was due to its reliance on a limited feature set (e.g., color, size, shape), making it susceptible to real-world variations. Based on the detailed classes derived from this analysis, this study trained a robust Deep Neural Network (DNN) using a comprehensive and well-structured training dataset. The resulting DNN-based tracker demonstrated significantly higher recognition rates and greater robustness across diverse environments compared to the legacy system. These results demonstrate that class refinement and data-driven error analysis significantly contribute to enhancing deep learning model performance.