Multiscale search using probabilistic quadtrees

We propose a novel framework to search for a static target using a multiscale representation. The algorithm we present is appropriate when the target detection sensor trades off accuracy versus covered area, e.g., when a UAV can fly and sense at different elevations. A structure based on quadtrees is used to propagate a posterior about the target location using a variable resolution representation that is dynamically refined in regions associated with higher probability of target presence. Probabilities are updated using a Bayesian approach accounting for erroneous sensor readings in the form of false positives and missed detections. The model we propose is coupled with a search and decision algorithm that determines where to sense next and with which accuracy. The search algorithm is based on an objective function accounting for both probability of detection and motion costs, thus aiming to minimize traveled distances while trying to localize the target. The paper is concluded with simulation results showing our approach outperforms commonly used methods based on uniform resolution grids.