Controlled sensing for classification using image-based sensor networks

Consider the problem of controlling a network of surveillance sensors that are capable of selecting which areas to observe and which modes to observe these areas. We study the problem of controlling the observations of these sensors adaptively in order to classify accurately a collection of objects using information on their observed features. Our proposed approach is modeling objects as templates of 3-D features, and modeling sensors as observing features of individual objects, subject to degradation by noise, obscuration, missed detections and background clutter. We formulate the classification problem using a statistical framework based on random feature sets, and present sequential inferencing algorithms for fusion of information from multiple sensor measurements. Using this framework, we focus on developing controlled sensing approaches that lead to efficient real-time algorithms for problems involving multiple sensors and objects. We develop approaches based on the use of information theoretic bounds that can be computed off-line for computing surrogate performance objectives, as well as finite parameterizations of collected information that are amenable to fast real-time computation. The latter approaches are also suitable for dynamic control strategy design using Partially Observed Markov Decision Process models. We compare the performance of the different approaches on a complex simulated scenario. Our results indicate that discrete measurement approximations have improved performance over the use of information theoretic bounds, and have performance that is statistically equivalent to more complex on-line optimization approaches, while reducing on-line computation requirements by over 5 orders of magnitude.