Development of a human computer interface based on hand gestures

As described up to now, the matching process works assuming that users always perform gestures, standing in the same position (i.e. at the same location and distance respect to the capturing camera) that was adopted when video models were captured. This is a very limiting assumption, as one of the major benefits of this way of interacting should be the possibility of a user to control the system, just standing in the field of view of the camera and without using any kind of device. If, in order to interact, the user must stand in a precise location, this benefit disappears.

In order to make trajectory features indepent from the user location, position and velocity must be normalized respect to something that moves together with the user, like the user’s face. If the centroid of the face location becomes the point (0, 0) of the euclidean plane in which hands moves, and the width and height become its unit respectively along x and y axes, then the trajectory gets invariant respect to translation and scale of the user. To detect the face bounding box, a standard face detector like the Viola-Jones can be adopted.

One of the constraints imposed on models is that their represented gestures must be performed, during video capturing, with just one hand; it is straightforward that in this way it is not possible to recognize gestures performed with more then one hand. Actually this problem can be simply bypassed if hands are tracked with the method described in section 2.3, and features extracted from each hand are processed independently by the spotting module. If these conditions are applied, then it is possible to spot at the same time, different gestures from each hand; thus gesture performed by more than one hand can be modeled as a combination of gestures spotted independently by each hand. During this work the spotting of gestures performed by more than one hand has not been implemented.