Journal of Computer Applications ›› 2020, Vol. 40 ›› Issue (9): 2507-2513.DOI: 10.11772/j.issn.1001-9081.2020010019

• Artificial intelligence • Previous Articles     Next Articles

Clustering relational network for group activity recognition

RONG Wei, JIANG Zheyuan, XIE Zhao, WU Kewei   

  1. College of Computer Science and Information Engineering, Hefei University of Technology, Hefei Anhui 230601, China
  • Received:2020-01-14 Revised:2020-04-17 Online:2020-09-10 Published:2020-04-28
  • Supported by:
    This work is partially supported by the Natural Science Foundation of Anhui Province (1808085MF168).


戎炜, 蒋哲远, 谢昭, 吴克伟   

  1. 合肥工业大学 计算机与信息学院, 合肥 230601
  • 通讯作者: 蒋哲远
  • 作者简介:戎炜(1994-),男,安徽合肥人,硕士研究生,CCF会员,主要研究方向:图像处理、深度学习;蒋哲远(1965-),男,安徽巢湖人,副研究员,博士,CCF高级会员,主要研究方向:软件理论与智能、面向服务软件工程;谢昭(1980-),男,安徽合肥人,副研究员,博士,主要研究方向:计算机视觉、图像理解;吴克伟(1984-),男,安徽合肥人,副研究员,博士,主要研究方向:计算机视觉、图像理解。
  • 基金资助:

Abstract: The current group behavior recognition method do not make full use of the group relational information, so that the group recognition accuracy cannot be effectively improved. Therefore, a deep neural network model based on the hierarchical relational module of Affinity Propagation (AP) algorithm was proposed, named Clustering Relational Network (CRN). First, Convolutional Neural Network (CNN) was used to extract scene features, and the regional feature clustering was used to extract person features in the scene. Second, the hierarchical relational network module of AP was adopted to extract group relational information. Finally, the individual feature sequences and group relational information were fused by Long Short-Term Memory (LSTM) network, and the final group recognition result was obtained. Compared with the Multi-Stream Convolutional Neural Network (MSCNN), CRN has the recognition accuracy improved by 5.39 and 3.33 percentage points on Volleyball dataset and Collective Activity dataset, respectively. Compared with the Confidence-Energy Recurrent Network (CERN), CRN has the recognition accuracy improved by 8.70 and 3.14 percentage points on Volleyball dataset and Collective dataset, respectively. Experimental results show that CRN has higher recognition accuracy in the group behavior recognition tasks.

Key words: group behavior recognition, Clustering Relational Network (CRN), group relational information, Affinity Propagation (AP) algorithm, Long Short-Term Memory (LSTM) network

摘要: 目前群组行为识别方法没有充分利用群组关联信息而导致群组识别精度无法有效提升,针对这个问题,提出了基于近邻传播算法(AP)的层次关联模块的深度神经网络模型,命名为聚类关联网络(CRN)。首先,利用卷积神经网络(CNN)提取场景特征,再利用区域特征聚集提取场景中的人物特征。然后,利用AP的层次关联网络模块提取群组关联信息。最后,利用长短期记忆网络(LSTM)融合个体特征序列与群组关联信息,并得到最终的群组识别结果。与多流卷积神经网络(MSCNN)方法相比,CRN方法在Volleyball数据集与Collective Activity数据集上的识别准确率分别提升了5.39与3.33个百分点。与置信度能量循环网络(CERN)方法相比,CRN方法在Volleyball数据集与Collective Activity数据集上的识别准确率分别提升了8.7与3.14个百分点。实验结果表明,CRN方法在群体行为识别任务中拥有更高的识别准确精度。

关键词: 群组行为识别, 聚类关联网络, 群组关联信息, 近邻传播算法, 长短时记忆网络

CLC Number: