Date of Publication :17th May 2024

Abstract:As a part of the AI revolution, Activity Recognition has found numerous applications over the decade. Recognition of cricket shots is one of the less researched domains, however, it has profound applications. Instant classification of cricket shots can be of immense help in training videos, performance analysis of the players or the game itself, and also in enriching the watching experience of the sport. This paper addresses issues involved in manual cricket shot classification, to achieve a state of automated shot recognition in the game. The manual techniques used for shot recognition are time-consuming and require expert assistance at all levels of the sport. Instant recognition of the shots avoids expert involvement and translates into a reduction of time and effort. Furthermore, its application has the potential to increase grassroots engagement and improve the game's analytical component. The success of deep learning algorithms with video and image data is one of the motivations for exploring similar algorithms for cricket shot classification. The suggested methodology uses several deep learning algorithms best suited for activity recognition and compares their performance. Specifically, the study delves deeply into CNN + RNNs, Attention Networks, and Vision Transformer, a specific type of CNN, utilizing temporal and spatial information to improve classification results. The experiments showed exceptional accuracies of 99.19% for Attention Networks, 99.2% for CNN + RNNs and 98.9% for ViTs on the PES dataset. The results obtained for the reference dataset had no significant difference to the PES dataset for ANN and CNN+RNN while slight difference for ViT model, validating our tested model. This technology adds a new level of knowledge by facilitating accurate shot detection and contextual relevance, which helps decision- making processes related to shot selection in all formats of the game which can be extended to other sports as well.

Reference :