多尺度子群体交互关系下的群体行为识别方法  

Group Activity Recognition under Multi-scale Sub-group Interaction Relationships

作　　者：朱丽萍 吴祀霖[1,2] 陈晓禾 李承阳 朱凯杰 ZHU Liping;WU Silin;CHEN Xiaohe;LI Chengyang;ZHU Kaijie(Beijing Key Laboratory of Petroleum Data Mining,Beijing 102249,China;College of Information Science and Engineering/College of Artificial Intelligence,China University of Petroleum(Beijing),Beijing 102249,China;School of Computer Science,Peking University,Beijing 100871,China)

机构地区：[1]石油数据挖掘北京市重点实验室,北京102249 [2]中国石油大学(北京)信息科学与工程学院/人工智能学院,北京102249 [3]北京大学计算机学院,北京100871

出　　处：《电子与信息学报》2024年第5期2228-2236,共9页Journal of Electronics & Information Technology

基　　金：北京市自然科学基金(L233002);中国石油科技创新基金(2022DQ02-0609)。

摘　　要：群体行为识别旨在识别包含多个个体的群体行为。在现实场景中,群体行为通常可以被视为从群体到子群体再到个人的层次结构。然而,以前的群体行为识别方法主要侧重于对个体之间的关系进行建模,缺乏对子群体之间关系的深度研究。从该角度出发,该文提出一种基于多尺度子群体交互关系(MSIR)的多层次群体行为识别框架。除对个体关系进行建模外,重点关注了子群体之间的多尺度交互特征。具体优化如下:设计子群体划分模块,通过个体外观特征和其空间位置来聚合可能存在关联的个体,再进一步利用语义信息动态地生成不同尺度大小的子群体;设计子群体交互特征提取模块,通过构建不同子群体之间的交互矩阵以及图神经网络的关系推理能力,提取更具判别力的子群体特征。实验结果表明,与现有12种方法在排球数据集和集体活动数据集这两个群体行为识别基准数据集上对比,该文方法都取得最好的性能结果。作为一个易于扩展和优化的群体行为识别框架,该算法在不同数据集上都具有较好的泛化能力。Group activity recognition aims to identify behaviors involving multiple individuals.In real-world applications,group behavior is often treated as a hierarchical structure,which consists group,subgroups and individuals.Previous researches have been focused on modeling relationships between individuals,without indepth relationship analysis between subgroups.Therefore,a novel hierarchical group activity recognition framework based on Multi-scale Sub-group Interaction Relationships(MSIR)is proposed,and an innovative multi-scale interaction features extraction method between subgroups is presented as specified below.A subgroup division module is implemented.It aggregates individuals with potential correlations based on their appearance features and spatial positions,then dynamically generates subgroups of different scales using semantic information.A sub-group interactive feature extraction module is developed to extract more discriminative subgroup features.It constructs interaction matrices between different subgroups and leverages the relational reasoning capabilities of graph neural networks.Compared with existing twelve methods on benchmark datasets for group behavior recognition,including volleyball and collective activity datasets,the methodology of this paper demonstrates superior performance.This research presents an easily extendable and adaptable group activity recognition framework,exhibiting strong generalization capabilities across different datasets.

关 键 词：行为识别 群体行为 子群体划分 关系推理 

分 类 号：TN911.7[电子电信—通信与信息系统] TP391.4[电子电信—信息与通信工程]