CricPose360: A Comprehensive Dataset for Real-Time Pose-Enhanced Cricket Shot Classification Using YOLOv8

Cricket shot classification is challenging due to the subtle differences in actions and dynamic environments. Accurate recognition is vital for context-based advertising, sensor-based commentary systems, and coaching assistants. In this study, we propose an approach for real-time, pose-enhanced shot classification and introduce CricPose360, a dataset with 3,030 cricket batting videos covering 10 different shots. Leveraging the YOLOv8 architecture, we implemented a two-step process to significantly boost classification accuracy. First, the YOLOv8l-pose model was employed to generate skeleton keypoint annotations from the raw videos, enriching the spatial and temporal features crucial for accurate shot recognition. These annotated videos were then used to extract 32 frames per video, which were fed into the YOLOv8l-cls model for classification. The incorporation of pose-enhanced data led to an accuracy increase of approximately 3%, achieving a final classification accuracy of 92.21% and f1-score of 92.29%. This proposed approach underscores the effectiveness of pose-based data augmentation in cricket shot classification, setting a new benchmark in sports analytics and providing a valuable dataset for future research.