A multiple perspective spectral approach to object detection

Abstract

Many applications for detection of objects such as video analysis require that candidate objects be observed over a range of perspectives in 3 dimensional space. As a result we must have a robust model and detection process for these objects in order to accurately detect them through a range of geometric transformations. In order to keep our detection process computationally efficient, we use a compact multiresolution model to represent the range of geometric transformations possible in the object to be detected. Additionally, we form an integrated likelihood ratio detection statistic to optimize the detection performance over the entire space of targets being examined. To demonstrate the performance of this algorithm we apply our results to a compressed video sequence and show the improvement of our integrated three dimensional model as a function of model order.