Multiple-Instance Support Vector Machine Based on a New Local Feature of Hierarchical Weighted Spatio-Temporal Interest Points

Abstract

Human action recognition is a hot research topic. However, in actual scene such as house intelligent monitoring, the background is disordered, many external factors harden the automatic recognition of human action. In this paper, we mainly paid attention to finding a feature to describe human actions efficiently and meanwhile deal well with intra-class and inter-class changes of human bodies, and also solve the problems that external factors cause. Thus, We proposed a new kind of feature, the Local Feature of Hierarchical Weighted Spatio-Temporal Interest Points, which fused different features in a specific way. To more accurately classify the presented features, based on Support Vector Machine, we introduced a new Multiple Instance Learning algorithm, forming the Multiple-Instance Support Vector Machine. Finally, we validated on the KTH public dataset and tested on the captured family activity video dataset. And we got a higher accuracy for human action recognition in home environment.