Human action classification using histogram-based discriminative embedding

In order to have a rich representation for human action, we propose to combine two complementary features so that a human posture can be characterized in more details. In particular, the distance signal feature and the width feature are combined in an effective way to enhance each other's discriminating capability. The resulting feature vector is quantized into mid-level features using k-means clustering. In the mid-level feature space, we apply the nonparametric embedding method to construct a compact yet discriminative subspace model. We have conducted a series of experiments on the Weizmann dataset to validate the proposed scheme. Compared with the existing approaches, our method can achieve high recognition accuracy while having a reduced computational complexity in classification stage.