Abstract: This thesis introduces a novel method for sequentially accumulating evidence as it pertains to an active observer seeking to identify an object in a known environment. First, a probabilistic framework is developed, based on a generalized inverse theory, where assertions are represented by conditional probability density functions. In order to resolve ambiguous assertions from single view measurements, a sequential recognition strategy is developed in which evidence is accumulated over successive viewpoints until a definitive assertion can be made. The main contribution of the thesis is a strategy for conditioning the inference and the measurement processes with feedback from prior information.
The problem of interest is that of model-based recognition, where the task is to identify an unknown model from a database of known objects on the basis of parameter estimates. The robustness of the algorithm is illustrated through its application to two very different domains: (1) recognition of 3-D parametric models estimated directly from laser rangefinder data, (2) recognition of objects based on signatures extracted from optical flow images that they generate as they move with respect to a camera. The latter approach is completely novel and presents a major contribution to the field. Experimental results verify the strength of the approach at overcoming difficulties encountered in both contexts, as rapid convergence to the correct solution occurs in most cases.
With this framework in place, it is further shown how an active recognition strategy could be built by using entropy maps to guide an active observer along an optimal trajectory for confidently inferring object identity and pose, while minimizing the amount of data that must be gathered. Specifically, these maps are used to encode prior knowledge about the discriminability of objects as a function of viewing position. The thesis applies the strategy to the context of recognition based on optical flow signatures, and shows how a gaze-planning strategy can be formulated by using entropy minimization as a basis for choosing a next best view. Experimental results are presented which show the strategy's effectiveness at converging to the optimal solution in a minimum short number of steps.