In the controlled active vision setting, where the output from a visual processing algorithm is to be used, there will naturally be a concern with the error or uncertainty associated to the visual processing algorithm. The standard approach to cleaning sensor noise is to add a filter to the sensor output. When vision is used in the feedback loop, how to properly filter the signal and where becomes a concern. Should the filtering be prior to the algorithm (i.e., of the image), after the algorithm (i.e., of the extracted state information), or of the actual algorithm itself?

Our, and other's earlier, research has shown that cleaning the image signal itself can destroy important statistical information in the image required by the vision algorithms, thereby degrading performance. Filtering the feedback signal will improve its high-frequency content, but does not always lead to the expected improvement when used as feedback. The filtering of the feedback signal may not respect the processing error nor the dynamics of the underlying image feature being used.

Rather, we argue here and have shown empirically in one case that a better approach would be to actually incorporate a filter into the algorithm. Thus, the research here looks into creating filtering methods fo computer vision algorithms and their internal states. In particular, we focus on segmentation-based tracking algorithms which are integral parts of a trackpoint maintenance system. Both the Bayesian segmentation and geometric active contour algorithms are studied in order to arrive at (near) optimal filtering strategies. We also look into related concepts, such as uncertainty estimation and adaptation, that serve to modify online the parameters of the filtering algorithms.

This project demonstrates and quantifies the performance improvements achieved through the use of an observer in concert with standard target tracking models. Tracking performance is measured against ground truth using the indicated location and the delineated boundary of the track target.