Tracking in Action Space

Abstract

The recognition of human actions such as pointing at objects (“ Give me that...” ) is difficult because they ought to be recognized independent of scene parameters such as viewing direction. Furthermore, the parameters of the action, such as pointing direction, are important pieces of information. One common way to achieve recognition is by using 3D human body tracking followed by action recognition based on the captured tracking data. General 3D body tracking is, however, still a difficult problem. In this paper, we are looking at human body tracking for action recognition from a context-driven perspective. Instead of the space of human body poses, we consider the space of possible actions of a given context and argue that 3D body tracking reduces to action tracking in the parameter space in which the actions live. This reduces the high-dimensional problem to a low-dimensional one. In our approach, we use parametric hidden Markov models to represent parametric movements; particle filtering is used to track in the space of action parameters. Our approach is content with monocular video data and we demonstrate its effectiveness on synthetic and on real image sequences. In the experiments we focus on human arm movements.