基于本征维数和置信度的行为序列分割  

Time-sequential Activity Segmentation Based on the Intrinsic Dimension and the Confidence

作　　者：熊心雨[1] 潘伟[1] 唐超[1] 

机构地区：[1]厦门大学信息科学与技术学院,福建省仿脑智能系统重点实验室,福建厦门361005

出　　处：《厦门大学学报（自然科学版）》2013年第4期479-485,共7页Journal of Xiamen University：Natural Science

基　　金：国家自然科学基金项目(60975084)

摘　　要：人机交互研究领域中行为分析与识别是当前研究的一个热点,行为序列分割是行为分析与识别的基础.鉴于强度摄像机视频在进行行为分割时对光线、视角变化过于敏感,提出了一种由深度视频提取的骨架信息,基于本征维数与置信度二次判断的无监督行为序列分割算法.首先,通过Kinect跟踪人体20个骨骼关节点数据,获得视频中人的姿态,通过提取关节点极坐标位置信息来描述行为特征;然后通过奇异值分解(sigular value decomposition,SVD)估计行为序列的本征维数,确定数据对应的低维流形,通过检测特征数据在该流形上投影误差的突变来找到分割帧,并对分割出来的行为序列进行类别标记.每找到一个分割帧就对当前标记类包含样本和当前标记类的前一类包含样本进行基于置信度的二次判断,找到前一类最优分割帧并初始化继续分割.最后采用随机森林模型对分割结果进行识别验证.实验结果表明采用本文算法可以明确分割出代表不同模式的行为片段.Behavior analysis and recognition in the current is a hot spot in the field of human-computer interaction,and the behavior sequence segmentation is the basis of behavior analysis and recognition. Rather than traditional human action recognition methods that are based on deploy single kind of descriptor to describe activities,a new unsupervised behavior segmentation algorithm,using human body skeleton video and being based on the intrinsic dimension and confidence is proposed in this paper. First of all, through the twen- ty joint polar coordinate data which is tracked by the Kinect,we can know the position of the person in the video. The behavior char- acteristics are described by the extractive joint polar coordinate position information; Secondly, the intrinsic dimensionality and the low-dimensional manifolds are determined using SVD,and the break of projecting error of activity sequence on the determinate mani- folds is detected as the segmentation point of the activity sequences. After finding the segmentation frame,marking the behavior se- quences belonging to the right class. Each time finding a segmentation frame, the secondary judgment based on the degree of confi- dence is used among samples belonging to the current marked class and samples belonging to the former marked class. Then find the optimal segmentation frame and initial to split. At last, using the random forest model to verify and identify the segmentation results. The experimental results show that using the algorithm proposed in this paper can segment the behavior sequences which can be on behalf of the different patterns.

关 键 词：行为分割 骨架模型 本征维数 置信度 随机森林 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]