1 Scene recognition using AlexNet to recognize significant events within cricket game footage
Tevin Moodley (University of Johannesburg), Dustin van der Haar (University of Johannesburg)
Computer Vision, Deep Learning, Cricket, Stroke Recognition
In the last decade, special attention has been made toward the automated analysis of human activity and other related fields. Cricket, as a research field, has of late received more attention due to its increased popularity. The cricket domain currently lacks datasets, specifically relating to cricket strokes. The limited datasets restrict the amount of research within the environment. In the study, this research paper proposes a scene recognition model to recognize frames with a cricket batsman. Two different classes are addressed, namely; the gameplay class and the stroke class. Two pipelines were evaluated; the first pipeline proposes the Support Vector Machine (SVM) algorithm, which undergoes data capturing, feature extraction using histogram of oriented gradients and lastly classification. The Support Vector Machine (SVM) model yielded an accuracy of 95.441%. The second pipeline is the AlexNet Convolutional Neural Network (CNN) architecture, which underwent data capturing, data augmentation that includes rescaling and shear zoom followed by feature extraction and classification using AlexNet. The AlexNet architecture performed exceptionally well, producing a model accuracy of 96.661%. The AlexNet pipeline is preferred over the Support Vector Machine pipeline for the domain. By recognizing a significant event, that is when a stroke and none stoke (gameplay) scene is recognized. The model is able to filter only relevant footage from large volumes of data, which is then later used for analysis. The research proves there is value in exploring deep-learning methods for scene recognition.