Entropy-Aware Activity Recognition for Sports: Case Study on Cricket

Abstract

In the digital era, activity recognition substantially impacts everyone’s day-to-day lives. Multiple people and their activities can be seen in the video, and their routines depend on video devices, specifically in sports. Video datasets in sports face several challenges, such as intra-class similarity, inter-class ambiguity, noise and motion transitions. The research proposed an entropy-aware deep learning method for activity recognition that integrates a pre-trained CNN with Density-Based Spatial Clustering of Applications with Noise (DBSCAN) for Cricket keyframe selection. A synthesized dataset of 25,000 annotated cricket clips is generated across five fine-grained activity classes. These keyframes are uniformly sampled, encoded into CNN embeddings, and clustered using DBSCAN. The chosen keyframes are then fine-tuned using a CNN to optimize using categorical cross-entropy loss. The extensive experiments show that the proposed method achieves an accuracy of 96.4% and provides a reliable 6% improvement over unclustered data. The findings show that frame selection using DBSCAN improves generalization across a variety of broadcasting conditions and provides a scalable framework for automated cricket video analysis, with real-time applications in highlight generation, intelligent sports surveillance, and player performance monitoring.