Cricket Video Summarization Using Temporal CNNs

Abstract

In this study, a new method for the summarization of very long cricket videos through the employment of an enriched deep learning approach is proposed and utilises the inherent feature of a Three-Dimensional Convolutional Neural Network (3D-CNN). Our methodology comprises several key stages as the first one is development and testing of Residual Network (ResNet) structure 3D-CNN for the identification and classification of significant events in cricket matches the second one is the annotation of the video clips divided into five classes of actions: fours, sixes, wickets, milestones and others and the third one is fine-tuning of the ResNet-based 3DCNN with the use of the annotated The model is expected to correctly detect important cricket events, and thus to assist in creating a highlight summary by keeps clips with fours, sixes, wickets, and milestones while discarding all unnecessary parts, to ensure we test the performance of our Summarization system, we conducted experiments by assessing the accuracy of the system after training the model on unseen cricket match videos got an average accuracy of 97%. Based on these results, it proves that our approach provides an efficient and accurate way to autonomously produce short and context-specific summaries for cricket matches using the 3D-CNN.