基于分阶段视觉共生矩阵序列的双人交互识别  

Human interaction recognition based on multi-stage framework and co-occurring visual matrix sequence

作　　者：姬晓飞[1] 左鑫孟 

机构地区：[1]沈阳航空航天大学自动化学院,辽宁沈阳110136

出　　处：《计算机工程与设计》2017年第9期2498-2503,共6页Computer Engineering and Design

基　　金：国家自然科学基金项目(61103123);辽宁省高等学校优秀人才支持计划基金项目(LJQ214018)

摘　　要：针对复杂场景下双人交互行为识别率低、算法复杂度高的问题,对双人交互行为识别框架和特征描述进行深入研究,提出一种分阶段视觉共生矩阵序列的双人交互识别方法。根据交互双方的连通性,将交互过程分为3个阶段,对开始阶段和结束阶段采用个体分割框架进行处理,用提出的视觉共生矩阵序列表示交互特征,对执行阶段用整体框架进行处理,提取HOG特征,利用HMM模型离线训练交互行为样本分类器,加权融合各阶段的识别概率,得到识别结果。利用UT-interaction数据库进行实验,实验结果表明,该方法可实现对交互行为的准确识别,满足实时需求。A human interaction recognition method based on the multi-stage framework and co-occurring visual matrix sequence was proposed to solve the problems, such as relatively low recognition accuracy and high computational cost, by studying the re-cognition framework and feature description. According to connectivity of interaction, interactive process was divided into three stages. Interaction recognition based on individual segmentation was used for start stage and finish stage, and co-occurring visual matrix sequence was extracted as their feature description, and execution stage with HOG feature was recognized based on the general framework. HMM model was adopted to offline train the sample classifier of interactive behavior, and the best identification result can be obtained through weighted fusion for multi-stage identification probability. Experimental results on the UT-interaction dataset show that the method achieves better recognition performance with simple implementation and satisfies the real time demand.

关 键 词：双人交互识别 分阶段框架 HOG特征 视觉共生矩阵序列 HMM模型 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]