Detection of military objects in LADAR images

This report describes different techniques for preprocessing, segmentation, and detection of vehicle sized objects in LADAR images. Five preprocessing strategies are presented; 1) Median filtering, 2) Two 1-D median filters in cascade, 3) Spoke median filter, 4) Donut filter, 5) Outlier detection and removal. The spoke median and donut filters were virtually worthless. The other filters worked equally well. The outlier detector removed outlers while perserving edges and small structures (and image noise). Concerning segmentation algorithms, we have implemented and tested four groups of region based algorithms and one group of edge based algorithms. Output from the segmentation is input to an object definition algorithm. Two strategies are proposed; one conventional agglomerative clustering approach, and one graph based approach. In essence, they both give the same results. Clusters with height, width, and length within predefined intervals are assumed to be possible objects. All algorithms are tested on real data of various vehicles in different scenes. It is difficult to draw any general conclusions. However, it seems that the region based algorithms perform better than the edge based ones. Among the region based strategies, those based on morphology or filtering operations perform well in most cases.