Human action recognition from silhouettes using manifold learning and MDA

In this paper, a method for human action recognition based on silhouette observations is presented. The methodology combines neighbourhood preserving embedding (NPE) based dimensionality reduction and multiple discriminant analysis (MDA) based action discrimination. Since depth is more robust to appearance variations and enables the spatial relationship to be explored easily for future multi-target interaction, a miniature stereo vision machine with three cameras is employed for generating high-resolution dense depth maps at video rate. The silhouettes of a moving person are extracted as the input features through fusing colour and depth information. Motivated by the observation that human activities are often located on a low-dimensional latent space, NPE is adopted to learn the intrinsic action manifold. When the class information is available, MDA is utilised to find a linear subspace which is optimal for discrimination. Experimental results show that our system can recognise human actions accurately with temporal, intra-and inter-person variations.