Deep learning for classification of cricket umpire postures

Among various multimedia resources on Internet, videos are the most popular among humans since they require less active brain power compared with static images, text and audio. With the exponential growth of videos available online, solutions for automatic content analysis in different contexts have attracted the attention of researchers. In diverse types of videos, generating highlights of sports videos is a highly considered segment by recent researchers. This research is focused on examining the existing deep learning architectures to analyze the feasibility of combining prevailing network models with classifiers while checking which combination gives the highest performance. A novel combination is presented to automatically detect the cricket umpire postures under five events (Six, No Ball, Out, Wide, None) using Convolutional Neural Networks (CNNs) and standard classifiers. The proposed method utilizes VGG16, ResNet50V2 and MobileNetV2 CNN architectures to extract the features and SVM and Naïve Bayes to classify the identified features into umpire postures. A new dataset has been constructed for this research which contains a total of 350 cricket umpire images. The results demonstrate that ResNet50V2 model in combination with SVM classifier gives the highest classification performance for the proposed dataset.