基于连续数据流的动态手势识别算法

Algorithm based on continuous data stream for dynamic gesture recognition

为识别用户做出的动态手势序列,基于数据手套采集的连续数据流,运用奇异值分解消除数据噪点,提取手势的特征信息,并利用关节弯曲的生理学特性与用户解耦合,将各种动作片段抽象成用户无关的手势模板,从而唯一定义手势特征并屏蔽不同用户的手势差异,再基于Hill Climbing思想把连续数据流分割成有序的动作序列,并按时序对所有片段在预先构造的层次树上实时搜索,根据欧式距离度量序列与手势模板的相似性.该算法对手势序列的分割准确,对多用户具有良好的适应性,其有效性在使用5DT数据手套搭建的两组动态手势识别的实验中得以验证.

For the purpose of recognizing the sequence of dynamic gesture made by operator, a method was presented based on continuous data streams sampled from data glove, which used singular value decompo- sition （SVD） to eliminating noise and extracting features. The characteristics of physiology about joint bend was applied making user-dependent information be culled. A set of gesture template which across different us- ers was set up. The template which gives a complete description of gesture＇ s feature and generalizes it is therefore user-independent. Based on Hill Climbing heuristic, these streams were separated into action se- quences, then a similarity measurement using Euclidian distance was adopted in real time between all seg- ments and templates on a hierarchy search tree built in advance. The sequences segmented by this method are accuracy and suitable for multi users. The effectiveness of this approach for identifying dynamic＇gesture was verified by two empirical experiments which using 5 DT data glove.