基于合作与竞争交互关系的组群行为识别算法

基于合作与竞争交互关系的组群行为识别算法

全文下载： 202205015.pdf

文章编号： 1672-6987（2022）05-0109-12； DOI： 10.16351/j.1672-6987.2022.05.015

王传旭， 林晓萌， 林国丞(青岛科技大学 信息科学技术学院，山东 青岛 266061)

摘要： 针对目前群组成员交互关系建模中缺乏清晰的高层语义描述，提出了一种基于“合作与竞争”交互关系建模的群组行为识别。首先利用弱监督方法实现对群组整体成员的自动分组和关系判断，即利用IAP(improved affinity propagation)算法实现对群组成员的自动分组，并利用Bert和GCN网络分别识别同一簇成员间存在的合作/竞争关系和不同簇间成员的合作/竞争关系；其次，利用半监督模块提供的身体部位信息以及部分帧中成员的信息进行迁移学习，从而识别每个成员的动作，将其动作标签信息进行编码后作为弱监督模块的补充；再次，利用深层聚合模型对上述两模块提取的语义和关系特征进行融合，最后利用softmax实现群组行为识别。为了验证本研究模型的有效性，选择在CAD和NBA数据集上进行实验，分别取得了942%和525%的效果优于当前较先进算法的识别效果。

关键词： 组群行为识别； 合作与竞争关系； 弱监督模块； 半监督模块

中图分类号： TP 301.6文献标志码： A

引用格式： 王传旭， 林晓萌， 林国丞. 基于合作与竞争交互关系的组群行为识别算法［J］. 青岛科技大学学报(自然科学版), 2022, 43(5): 109-120.

WANG Chuanxu, LIN Xiaomeng, LIN Guocheng. Group activity recognition algorithm based on the interaction between cooperation and competition［J］. Journal of Qingdao University of Science and Technology（Natural Science Edition), 2022， 43（5）： 109-120.

Group Activity Recognition Algorithm Based on the Interaction Between

Cooperation and Competition

WANG Chuanxu, LIN Xiaomeng, LIN Guocheng

(College of Information Science and Technology, Qingdao University of Science and Technology, Qingdao 266061, China)

Abstract: Aiming at the lack of clear high-level semantic description in current group member interaction modeling, we propose a group behavior recognition based on "cooperation and competition" interaction modeling. First, the weak supervision method is used to realize the dynamic grouping and relationship judgment of the overall members of the group, that is, the IAP (improved affinity propagation) algorithm is used to realize the automatic grouping of group members, and the Bert and GCN networks are used to identify the existence of the same cluster members. Cooperation/competition relationship and the cooperation/competition relationship between members of different clusters; secondly, use the body part information provided by the semi-supervised module and the information of members in some frames for migration learning, thereby identifying the actions of each member, and carrying out their action information After encoding, it serves as a supplement to the weak supervision module; again, the deep aggregation model is used to fuse the semantic and relational features extracted by the above two modules, and finally the Softmax is used to realize group behavior recognition. In order to verify the effectiveness of the model in this paper, experiments were conducted on the CAD and NBA datasets, and the results were 942% and 525% respectively.

Key words: group activity recognition; cooperation and competition; weak supervision model; semi-supervised model

收稿日期： 2021-09-26

基金项目： 国家自然科学基金项目(61672305).

作者简介： 王传旭(1968—), 男, 教授.