| This thesis presents a novel approach to active visual recognition whereby a sensor is guided towards the acquisition of a small set of observations from which to perform a recognition task accurately. The key contribution of the proposed approach is that the computational cost of observation selection is significantly lower than with current approaches presented in the literature. First, a probabilistic approach to the fusion of data obtained from multiple observations is presented. This approach works by recursively updating probabilities associated with each hypothesis, taking into account the different sources of uncertainty as well as the dependencies between the observed scene, the measured data and the observation parameters. A novel criterion for automatic selection of additional observations is then described. This criterion associates high utility with observations whose outcome predictably facilitates distinction between pairs of competing hypotheses. The algorithm is shown to have low complexity and lends itself to various simplifications. Therefore, it represents a much less expensive solution than techniques commonly used in the literature such as maximisation of mutual information or expected loss of entropy. The very general methodology presented is then applied to the open and difficult problem of object recognition and pose estimation. Rigourous experimentation of the proposed approach was conducted in two different contexts. (Abstract shortened by UMI.). |
