Detection of small moving objects in image sequences using multistage hypothesis testing

The detection of small, low-contrast, moving objects in a time sequence of digital images is addressed. Since object positions and velocities are unknown, a large number of candidate trajectories, organized into a tree-structure, are hypothesized at each pixel. At each `root´ image pixel, trajectory extensions are mapped to tree nodes. Pixels along a trajectory are tested sequentially for a shift in mean intensity using multistage hypothesis testing (MHT). The MHT is designed according to prespecified error probabilities. Exact, closed-form expressions for MHT test performance are derived and then applied to predicting the algorithm´s computation and memory requirements. Under a Gaussian white noise background assumption it is shown theoretically that over 4000 candidate trajectories per pixel are tested using an average of only 30 additions and threshold comparisons