Kernel PCA of HOG features for posture detection
Cheng Peng, Wanqing Li, Philip Ogunbona · 2009
Motivated by the non-linear manifold learning ability of the kernel principal component analysis (KPCA), we propose in this paper a method for detecting human postures from single images by employing KPCA to learn the manifold span of a set of HOG features that can effectively represent the postures. The main contribution of this paper is to apply the KPCA as a non-linear learning and open-set classification tool, which implicitly learns a smooth manifold from noisy data that scatter over the feature space. For a new instance of HOG feature, its distance to the manifold that is measured by its reconstruction error when mapping into the kernel space serves as a criterion for detection. And by combining with a newly developed KPCA approximation technique, the detector can achieve almost real-time speed with neglectable loss of performance. Experimental results have shown that the proposed method can achieve promising detection rate with relatively small size of positive training dataset.