A Deep Learning Based Approach for the Categorization of Actions in Cricket Videos

Cricket, a widely popular sport, has garnered significant attention, yet the classification of cricket actions remains underexplored in the research domain. This study aims to fill that gap by introducing CricShot13, an extended dataset that includes 3 new action categories-increasing the total to 13 distinct cricket actions-addressing the limitations of existing datasets in terms of scope and real-world applicability. The methodology combines ResNet50 for feature extraction and LSTM for sequence modeling to enhance the accuracy of cricket action classification. Keyframes are extracted from video sequences, and ResNet50 captures spatial features, while LSTM models the temporal relationships between frames. The computational complexity of the ResNet50 + LSTM model, with 512 hidden units, involves 28.90 million parameters, 7.74 Giga FLOPs, and 3.87 Giga MAC operations, indicating its efficiency in handling large-scale datasets. Our proposed approach achieves a classification accuracy of 90.04%. These results underscore the potential of this framework for applications in automated sports video analysis, coaching, and player performance assessment, contributing significantly to the field of computerized sports analysis.