Object recognition using an extended condensation filter

Abstract

This dissertation describes and tests a set of Condensation Filter extensions that fuse the pose estimates provided by recognition algorithms, while providing a pose localization mechanism to fuse similarity measures from classification algorithms — the traditional Condensation Filter can only accommodate pose estimates. The extensions also fuse the ad hoc information provided by low-level features to serve as heuristics and focus the target pose search. The ad hoc information is combined using the pose of the filter particle as a common reference point and mapping functions derived with Bayesian Learning. These extensions provide a more flexible framework for better representing and combining diverse sources of information than previously possible. Pose space is also examined more extensively where the conditional target pose probability is higher and more particles are present. This feature gives the Condensation Filter a new active role in directing the use of classification algorithms. The scaled circular error probable metric (CEP) is used to study the effect on object recognition of combining multiple recognition, classification, and low-level feature extraction algorithms using two different expert combination rules. To examine the relationship between discrimination, accuracy, and precision, receiver operating characteristic curve measurements are compared to the scaled CEP measurements. These results show little correlation between discrimination and accurate and precise localization of the target. The tests performed for this dissertation indicate the Extended Condensation Filter can reliably fuse recognition, classification, and raw feature data to perform more accurate and precise object recognition than is possible using each of the individual contributing algorithms.