Detecting Small, Moving Objects in Image Sequences Using Sequential Hypothesis Testing

A new algorithm is proposed for the solution of an important class of multidimensional detection problems: The detection of small, barely discernible, moving objects of unknown position and velocity in a sequence of digital images. A large number of candidate trajectories, organized into a tree structure, are hypothesized at each pixel in the sequence and tested sequentially for a shift in mean intensity. The practicality of the algorithm is facilitated by the use of multistage hypothesis testing (MHT) for simultaneous inference, as well as the existence of exact expressions for MHT test performance in Gaussian white noise (GWN). These expressions predict the algorithm’s computation and memory requirements, where it is shown theoretically that several orders of magnitude of processing are saved over a brute-force approach based on fixed sample-size tests. The algorithm is applied to real data by using a robust preprocessing procedure to eliminate background structure and transform the image sequence into a residual representation, modeled as GWN. Results are verified experimentally on a variety of video image sequences.