Optimal Positioning of Multiple Cameras for Object Recognition Using Cramér-Rao Lower Bound

In this paper the problem of active object recognition/pose estimation is investigated. The Principle Component Analysis is used to produce an observation vector from images captured simultaneously by multiple cameras from different view angles of an object belonging to a set of a priori known objects. Models of occlusion and sensor noise have been incorporated into a probabilistic model of sensor/object to increase the robustness of the recognition process with respect to such uncertainties. A recursive Bayesian state estimation problem is formulated to identify the object and estimate its pose by fusing the information obtained from the cameras at multiple steps. In order to enhance the quality of the estimates and to reduce the number of images taken, the positions of the cameras are controlled based on a statistical performance criterion, the Cramér-Rao Lower Bound (CRLB). Comparative Monte Carlo experiments conducted with a two-camera system demonstrate that the features of the proposed method, i.e. information fusion from multiple sources, active optimal sensor planing, and occlusion modelling are all highly effective for object classification/pose estimation in the presence of structured noise.