Multiphase action representation for online classification of motion capture data

Abstract

In this paper we introduce a novel, simple, and efficient method for human action recognition based on a multiphase representation of human motion. An action is considered as a finite state machine where each state represents a primitive motion called motion phase, which is simply a sequence of poses with predefined common features. Spatial-temporal and postural features introduced in previous work are redefined by using only 3D joint positions for features extraction and are extended by involving the relative movement of the body end-effectors as new features. We developed a framework for modelling a given motion in the proposed motion model, whereupon we used this framework to create a model database of 25 different actions. Using this database we conducted a number of experiments on data obtained from several sources as well as on distorted data. The results showed that the presented method has high accuracy and efficiency. Additionally, it can work offline and online in real time, and can be easily adapted to work on 2D data.