Human action recognition in 3D motion sequences

In this paper we propose a method for learning and recognizing human actions on dynamic binary volumetric (voxel-based) or 3D mesh movement data. The orientation of the human body in each 3D posture is estimated by detecting its feet and this information is used to orient all postures in a consistent manner. K-means is applied on the 3D postures space of the training data to discover characteristic movement patterns namely 3D dynemes. Subsequently, fuzzy vector quantization (FVQ) is utilized to represent each 3D posture in the 3D dynemes space and then information from all time instances is combined to represent the entire action sequence. Linear discriminant analysis (LDA) is then applied. The actual classification step utilizes support vector machines (SVM). Results on a 3D action database verified that the method can achieve good performance.