Using multiple sensors for reliable markerless identification through supervised learning

• Authors: Andrea Albarelli;Filippo Bergamasco;Augusto Celentano;Luca Cosmo;Andrea Torsello
• Venue: Machine Vision and Applications
• Year: 2013

In many interaction models involving an active surface, there is a need to identify the specific object that performs an action. This is the case, for instance, when interactive contents are selected through differently shaped physical objects, or when a two-way communication is sought as the result of a touch event. When the technological facility is based on image processing, fiducial markers become the weapon of choice in order to associate a tracked object to its identity. Such approach, however, requires a clear and unoccluded view of the marker itself, which is not always the case. We came across this kind of hurdle during the design of a very large multi-touch interactive table. In fact, the thickness of the glass and the printed surface, which were required for our system, produced both blurring and occlusion at a level such that markers were completely unreadable. To overcome these limitations we propose an identification approach based on SVM that exploits the correlation between the optical features of the blob, as seen by the camera, and the data coming from active sensors available on the physical object that interacts with the table. This way, the recognition has been cast into a classification problem that can be solved through a standard machine learning framework. The resulting approach seems to be general enough to be applied in most of the problems where disambiguation can be achieved through the comparison of partial data coming from multiple simultaneous sensor readings. Finally, an extensive experimental section assesses the reliability of the identification.