Categorization of actions in soccer videos using a CNN-RNN architecture

Abstract

The extraction of semantic information from soccer videos has several applications, such as contextual advertising, match summaries, and highlight extraction. Applications for analyzing soccer videos can be categorized into Action Detection, Player and/or Ball Tracking, and Game Analysis. A modified version of the SoccerNet-v2 Dataset is used as a
database to reduce the minimum computational power required. The task of Action Detection becomes challenging due to the overlap of actions and the various video capture conditions that include multiple angles, ads, and camera cuts. To overcome these challenges, a Convolutional Neural Network (CNN) and a Recurrent Neural Network (RNN) are used together to classify different lengths of soccer action videos. A pre-trained CNN, InceptionV3, is used for spatial
feature extraction, and a Gated Recurrent Unit (GRU) RNN is used for sequence recognition, which addresses temporal dependence and solves the problem of gradient disappearance. Finally, the Softmax layer assigns decimal probabilities to each class. A network configuration with four classifiable actions and an accuracy of 94% is achieved.