Human Action Recognition with Two-Level SVMs

Abstract

We propose a novel method for recognizing human actions from video sequences using multiclass support vector machines (SVMs). The proposed method enables the optimal combination of different space-and-time features. Our method is categorized as a deep model in which SVMs at the first-level are used to generate class probabilities based on each frame in the videos. At the second level, the SVM finally classifies actions by combining class probabilities of each video stream through the action time interval. Our twolevel SVM architecture can also combine different kinds of features efficiently and give better results than Bayes combination. We used two kinds of sparse features, i.e., optical flows and ST patches. Our experiments using the KTH dataset show that the two-level SVM architecture is effective for combining different kinds of features through action time intervals.