Target tracking in an urban warfare environment using particle filters

Particle filters have become a viable solution to the visual tracking problem as they are well suited for state estimation in the rich visual world where non-linear, non-Gaussian distributions are common and where information from various sensor sources must be fused. This paper introduces an adaptive particle filter framework in which the numbers of particles, fusion parameters, as well as filter parameters are updated during the filtering process. Then, a neural network is used to determine how well the filter is performing. Such a filter could be used within a hierarchical architecture for adversarial reasoning onboard an unmanned aerial vehicle executing reconnaissance and surveillance missions in an urban warfare environment. The innovative features of the tracking methodology include an adaptation mechanism used to focus on those cues that optimize the particle population and, therefore, reduce the computational burden, an automatic initialization technique that circumvents the need for an ad hoc selection of the initial particles, and the neural network classifier for performance assessment purposes. Initial implementation and testing of video images obtained from a moving vehicle and using color and motion cues suggest the efficacy of the proposed approach