Event Detection from Cricket Videos Using Video-Based CNN Classification of Umpire Signals

Cricket is a long duration game. It can last for several hours or even days. Due to cricket’s lengthy nature, it becomes indispensable for viewers to have the option of watching selected interesting events from a cricket match. As cricket videos are unscripted in nature, detecting key events from cricket videos is a big challenge. In this paper, we propose a method to detect 5 key events, namely FOUR, SIX, OUT, NO BALL and WIDE from cricket videos by recognizing the umpire’s signal. Analysis of umpire signals is performed using a pretrained Convolutional Neural Network (CNN) architecture, namely I3D. A new dataset, containing 504 videos and 2000 images from cricket matches, is introduced in this work for the detection of key events in cricket. For the tasks of umpire frame detection and umpire signal recognition, test accuracies of 97.76% and 86.14% are achieved respectively. Employing the proposed framework on official cricket match videos, we achieve precision, recall and F1 scores of 95.23%, 86.95% and 90.90% respectively. Furthermore, by evaluating performances of both proposed video-based approach and existing image-based approach on novel cricket videos, we can affirm that the proposed approach outperforms the image-based one.