Fast adaptive sensor management for feature-based classification

Modern surveillance systems often employ multiple steerable sensors that are capable of collecting information on selected objects in their environment. These sensors must coordinate their observation strategies to maximize the information collected by their measurements in order to accurately estimate the states of objects of interest. Adaptive sensor management consists of determining sensor measure- ment strategies that exploit previously collected information to determine current sensing actions, which are applied to the objects of interest.;In this dissertation, we focus on sensor management applied to the problem of classifying objects using small teams of sensors with limited capacity. Previous work on this sub ject assumes that sensors provide conditionally independent estimates of object type, which is unrealistic for modern image-based sensors. We address this shortcoming by developing a new mathematical theory for sensor management that models sensors as providing observations of features as opposed to object types. In our theory, objects are modeled as spatially related collections of features, characterized by object type and pose; sensors measure noisy pro jections of these features sub ject to degradation by noise, obscuration, missed detections and added background clutter. The performance of a classier that is based on a set of observed features depends on the accuracy with which features are measured, and how well the measured features are able to discriminate between object types. A key step in our theory is the processing of past measurements to provide supporting information for selecting sensing actions that improve classication accuracy. We develop a statistical framework based on random sets to characterize the relationship between observed features and object types, and obtain recursive estimates of probability distributions over object pose and type, using a generalized maximum likelihood approach.;The efficacy of sensor management algorithms depends of their ability to predict the value of information collected by potential measurements. A common approach to this problem is based on computation-intensive simulations of potential measurements and associated inference to evaluate expected values of information-theoretic metrics such as entropy. We develop a novel approach that combines an o?-line computation of apriori value of measurements using Bhattacharyya distances, and real-time estimates of object type and pose generated from past measurements, to generate a prediction of measurement value for sensor management. This value is based on a lower bound on the probability of classication error. We develop assignment algorithms to compute sensor management strategies to minimize this bound. The resulting sensor management algorithms are capable of solving problems involving a large numbers of objects in real-time.;To evaluate our proposed sensor management algorithm, we build synthetic 3-D models of object classes and simulate sensors as extracting features from 2-D pro jections of these models. We compare the performance of our real-time sensor management algorithm with other information-theory approaches that use measurement simulations. Our real-time algorithms achieve comparable classication accuracy, while requiring nearly three orders of magnitude less computation. Our results establish the feasibility of a practical, scalable and accurate approach for the real- time management of a team of sensors.