融合多特征的行为识别方法研究

Research on action recognition via multi-feature fusion

基于时空特征的方法是行为识别的主流方法,已经有许多研究学者提出了多种局部时空特征。然而,不同的局部特征所反映的行为信息的侧重点并不一样。通过引入集成学习的方法,对多种特征在分类器层次上进行融合,使得多种特征能够优势互补,从而增强了特征的描述能力,为构建出高效、稳定的行为识别分类器提供了保证。经仿真实验验证,所提出的方法是鲁棒和有效的。

The approach based on the local spatial-temporal features has emerged to be the mainstream method in action recognition area. And various descriptors of local spatial-temporal feature have been presented by researchers. However,different local features may reflect different emphasis of human activity. In this paper, the ensemble learning methods are introduced to perform a late fusion of multiple features so as to enhance the expressing ability of the local features. By the fusion of features, a more effective and robust action classifier can be built up. And the experimental results demonstrate the robustness and effectiveness of the proposed method.