• 算法设计与分析 •

### 基于交互关系分组建模融合的组群行为识别算法

1. （青岛科技大学信息科学技术学院,山东青岛266061）
• 出版日期:2022-01-24 发布日期:2022-01-24
• 作者简介:王传旭(1968—),男,山东邹城人,教授,硕士生导师,研究方向:计算机视觉,E-mail: Wangchuanxu_qd@163.com; 通信作者：刘冉(1995—),男,硕士研究生,研究方向:计算机视觉与模式识别,E-mail: liuran2016@163.com。
• 基金资助:
国家自然科学基金资助项目(61672305， 61802217)

### Group Activity Recognition Algorithm Based on Interaction Relationship Grouping Modeling Fusion

1. （College of Information Science and Technology, Qingdao University of Science and Technology, Qingdao 266061, China）
• Online:2022-01-24 Published:2022-01-24

Abstract: The modeling of interaction relationship between group members is the core technology of group activity recognition. High complexity and information redundancy in relational reasoning are tough problems in complex scenarios when modeling its group interactions. In order to solve these problems, we propose a model of grouping interactive relation. Firstly, CNN and RoIAlign are used to extract the scene information and personal information as initial features in each frame, and the whole group is divided into two subgroups by the personal spatial coordinates (For example, in the Volleyball data set, the X coordinates of participants’bounding boxes are used to rank, then, everyone set is set up an ordinal ID and 12 people are divided into two group from left to right). Secondly, the two local groups and the global scene groups are divided, the Graph Convolutional Network （GCN） is used to deduce their interaction relationship respectively, and the key persons in each group are determined. Then, we can regard global relationship features as the real value, and merge the characteristics of local relation of two groups as predicted value. In order to match the key figures of two groups with key figures from the whole group successfully, the cross-entropy loss function is built between the two and feedback to optimize the upper-level group GCN interaction relationship network. Next, with the information of key figures in the global interaction relationship as a guide, the key figures in the two subgroups are matched respectively. After successful matching, the matched key figures in the two subgroups are taken as the target nodes to establish a relationship graph between these two subgroups, and then it is deduced by GCN. Finally, the initial features are fused with intergroup and global interaction characteristics respectively to obtain two group behavior branches, and the final recognition result is obtained through decision fusion. The experiment shows that the accuracy is 93.1% on Volleyball data set and the accuracy is 48.1% on NBA data set.