nextuppreviouscontents
Next:Autonomous Physics-Based Colour Learning Up:Computer Vision Previous:Recognizing Objects by Accumulating

Viewpoint Selection by Navigation through Entropy Maps

T. Arbel, F.P. Ferrie

This project combines our work in sequential object recognition with active vision to create a process which can unambiguously recognize an object in a minimum number of views. Applications of this research include navigation systems that determine their positions relative to fixed landmarks in the environment and machine vision systems that must determine the position and orientation of known objects in potentially unstructured environments. The latter application is characteristic of work we have done with Hymarc Ltd. for the Canadian Space Agency. Our approach is based on a sequential recognition strategy in which object hypotheses are represented as conditional probability density functions. A Bayesian accumulation (chaining) method is used to accumulate evidence for the different hypotheses as new data are gathered on line. Selection of viewpoints is accomplished using an active vision approach that selects on the basis of minimizing ambiguity of recognition. The same off-line training process that is used to determine the prior conditional probability density functions used for recognition is also used to construct entropy maps relating ambiguity as a function of viewing position. These are effectively used by the active vision process to plan gaze. Preliminary experiments have shown that the combined use of sequential recognition and gaze planning greatly enhance the robustness of appearance-based methods. We are now investigating applications of object recognition using optical flow measurements as the basic input to the system.
 


Annual Report

Fri Nov 26 23:00:32 GMT 1999