アイテム詳細

要旨

In this chapter, we present a model-free pose estimation algorithm to estimate the relative pose of a rigid object. In the context of human motion, a rigid object can be either a limb, the head, or the back. In most pose estimation algorithms, the object of interest covers a large image area. We focus on pose estimation of objects covering a field of view of less than 5$^\circ$\ by 5$^\circ$\ using stereo vision. With this new algorithm suitable for small objects, we investigate the effect of the object size on the pose accuracy. In addition, we introduce an object tracking technique that is insensitive to partial occlusion. We are particularly interested in human motion in this context focusing on crash test dummies. The main application for this method is the analysis of crash video sequences. For a human motion capture system, a connection of the various limbs can be done in an additional step. The ultimate goal is to fully obtain the motion of crash test dummies in a vehicle crash. This would give information on which body part is exposed to what kind of forces and rotational forces could be determined as well. Knowing all this, car manufacturers can optimize the passive safety components to reduce forces on the dummy and ultimately on the real vehicle passengers. Since camera images for crash videos contain the whole crash vehicle, the size of the crash test dummies is relatively small in our experiments. For these experiments, mostly high-speed cameras with high resolution are used. However, the method described here easily extends to real-time robotics applications with smaller VGA-size images, where relative pose estimation is needed, {e.g.}\ for manipulator control.