Deep CNN Based Data-Driven Recognition of Cricket Batting Shots

Cricket is one of the most played and watched sports, specially in the South Asian region. This paper deals with identifying and categorizing various batting shots from cricket videos. Proposed method is based on deep convolution neural networks. Results have been evaluated for both 2D convolution followed by recurrent network for processing sequence of video frames and 3D convolution network for capturing spatial and temporal features simultaneously. In order to train and evaluate models, dataset comprising of about 800 batting shot clips have been locally developed. Obtained models are able to recognize a shot being played with 90% accuracy. The distinction of such visually similar shots with this high accuracy is novel in literature and indicates the high implications of modern AI and deep learning in applications for detecting various cricket activities as well as for decision making purposes. The prepared dataset will be made publicly available for research community.