摘要
以时空兴趣点特征和支持向量机(SVM)分类器识别方法为基本算法,在广泛使用的公开动作数据集KTH上,从不同角度考察评估策略对动作识别算法性能的影响。实验表明,当采用不同的交叉实验方法时,算法性能的波动最大达到10.5%,而不同数据集划分方法对算法性能的影响则达到11.87%。因此,通过量化分析得出的结论,可以充分地比较现有算法的真实差异,并为设计合理的评估策略提供参考。
Action recognition is a hot research topic,but the performance assessment strategies of algorithms have not had an accepted practice.In this paper,we adopt spatio-temporal features and support vector machine（SVM） model as our action recognition algorithm,and then well assess the effect of different assessment strategies on our action recognition algorithm in widely used public dataset KTH.Experimental results show that when different cross-experimental methods are employed,the performance fluctuation of algorithms reaches 10.5%.And when different division methods for KTH datasets are used,the performance fluctuation of algorithms gets 11.87%.Thus,according to conclusions in this paper,we can find the real difference among existing algorithms,and supply the reference for designing reasonable assessment strategy.
出处
《光电子．激光》
EI
CAS
CSCD
北大核心
2012年第6期1166-1172,共7页
Journal of Optoelectronics·Laser
基金
国家自然科学基金(90920001)
天津市科技支撑计划(10ZCKFGX00400)资助项目
关键词
时空特征
支持向量机(SVM)
动作识别
评估策略
spatio-temporal features
support vector machine（SVM）
action recognition
assessment strategy
