Ankhzaya JAMSRANDORJ Vanyi CHAO Yin May OO Mun, Kyung Ryoul Kim, Jinwook 2024-01-12T02:48:22Z 2024-01-12T02:48:22Z 2022-11-30 2022-11-11 https://pubs.kist.re.kr/handle/201004/76545 Most group activity recognition models focus mainly on spatio-temporal features from the players in sports games. Often they do not pay enough attention to the game object, which heavily affects not only individual action but also a group activity. We propose a new group activity recognition model for sports games that incorporates players’ motion information and game object positional information. The proposed method uses a transformer encoder for temporal feature extraction and a ’simple’ conventional convolutional neural network for extracting spatial features and fusing them with the relative ball position-embedded features. The experimental results show that our model achieved comparable results to state-of-the-art methods on the Volleyball dataset by using only one transformer encoder block and the ball position. English International Conference on Machine Learning and Intelligent Systems Ball position feature embedded Group Activity Recognition model for Team Sport Games Conference 10.3233/FAIA220435 1 The 4th International Conference on Machine Learning and Intelligent Systems (MLIS 2022) The 4th International Conference on Machine Learning and Intelligent Systems (MLIS 2022) KO 2022-11-08 Frontiers in Artificial Intelligence and Applications