基于多流架构与长短时记忆网络的组群行为识别方法研究  

Research on Group Behavior Recognition Method Based on Multi-Stream Architecture and Long Short-Term Memory Network

作　　者：王传旭[1] 胡小悦 孟唯佳 闫春娟 WANG Chuan-xu;HU Xiao-yue;MENG Wei-jia;YAN Chun-juan(Institute of Information Science and Technology,Qingdao University of Science and Technology,Qingdao,Shandong 266001,China)

机构地区：[1]青岛科技大学信息科学技术学院,山东青岛266001 [2]青岛科技大学信息科学技术学院通信教研室,山东青岛266001

出　　处：《电子学报》2020年第4期800-807,共8页Acta Electronica Sinica

基　　金：国家自然科学基金(No.61672305)。

摘　　要：提出一种基于多流架构与长短时记忆网络的上下文建模框架,旨在克服组群行为识别的两个难点,其一对复杂场景中多视觉线索进行信息融合;其二对情景人物进行建模,以获得长视频上下文关系.并且,对基于全局信息和基于局部信息的识别结果进行决策融合,判定最终组群行为属性.该算法在CAD1和CAD2上分别取得93.2%和95.7%平均识别率.This paper proposes a context modeling framework based on multi-stream architecture and LSTM,which aims to overcome two difficulties for group behavior recognition.One is to fuse information from multiple visual cues in complex scenes,the other is to model situational characters to get the long-term temporal context in the video.In addition,decision fusion is performed on the behavior recognition results based on global information and local information to determine the final group behavior attributes.The algorithm achieved 93.2%and 95.7%average recognition rates on CAD1 and CAD2 respectively.

关 键 词：组群行为识别 多视觉线索融合 交互上下文建模 全局-局部模型 长短时记忆网络 

分 类 号：TP301.6[自动化与计算机技术—计算机系统结构]