Ball detection for boccia game analysis

Abstract

The present article proposes the training, testing and comparison of two models for ball detection, taking into account its final implementation in a Boccia game analysis computer-vision algorithm, within the 'iBoccia' framework. The goal is to have a versatile and flexible algorithm towards different game environments. The selected ball detectors were a Histogram-of-Oriented-Gradients feature based Support Vector Machine (HOG-SVM) and a Convolutional Neural Network (CNN) based on a less complex implementation of the You Only Look Once model (Tiny-YOLO). Both detectors were evaluated offline and in real-time. The subsequent results showed that their performance was similar in both evaluations, however, Tiny-YOLO outperformed HOG-SVM by a small margin in all the used metrics. In real-time, both detectors achieved an accuracy of approximately 90%. Despite the high accuracy values, the detector requires further improvement because a single non-detection can influence the computer-vision algorithm's output, making the system unreliable.