Human action recognition by feature-reduced Gaussian process classification

Abstract

This paper presents a spectral analysis-based feature-reduced Gaussian Processes (GP) classification approach to recognition of articulated and deformable human actions from image sequences. Using Tensor Subspace Analysis (TSA), space–time human silhouettes extracted from action sequences are transformed to a low dimensional multivariate time series, from which structure-based statistical features are extracted to summarize the action properties. GP classification, based on spectrally reduced features, is then applied to learn and predict action categories. Experimental results on two real-world state-of-the-art datasets show that the GP classification outperforms a Support Vector Machine (SVM). In particular, spectral feature reduction can effectively eliminate the inconsistent features, while leaving performance undiminished. Moreover, compared with Automatic Relevance Determination (ARD), the spectral way for feature reduction is more efficient.