Cricket Umpire Intelligence: A CNN-SVM Framework for Action Detection
Ankita Suryavanshi, Shiva Mehta, Samir Rana, Kireet Joshi · 2024
This paper reports a novel approach that detects the umpire's actions using a combination of CNN and SVM networks. The focus is on recognizing five necessary umpire signals: The world of cricket is a world of its own. We see ‘Out,’ ‘Not Out,’ ‘Wide,’ ‘No Ball,’ and ‘Four’ among other inexplicable words. A dataset has been generated and annotated rigorously to measure the effectiveness of our approach. Afterwards, the extracted features are fed to a convolutional neural network (CNN) to identify the fraud patterns, and then SVM is used to classify them. The system's evaluation is based on different quality indicators that convey each class's accuracy and recall rates, as well as other quality metrics. The precision picketing ranges from 82.52% for Class 1 to 90.42% for Class 5, while the recall shots differ from 79.81% for Class 5 to 93.29% for Class 4. These statistics represent a model's immediate ability to affirm the activities of a specific person. The F -1-scores, which measure a balance between correct identification of the relevant details and recall of all details, are also provided. Class 4 comes on top among the classes, with participating students achieving the highest score (90.26%). As for the overall performance, the individual ratings are mixed with those averages, like macro, micro, and weighted. The macro average gives a score of 85.89%. The micro-average rating is a permanently high 85.84% for all the precision, recall, and F1-score measures. The F1 score after all the weighted adjustments is approximately 0.8579, which is led by the need to compensate for the class inequalities.