Multistatic ground-penetrating radar (GPR) signals can be imaged tomographically to produce 3-D distributions of image intensities. In the absence of objects of interest, these intensities can be considered to be estimates of clutter. These clutter intensities spatially vary over several orders of magnitude and vary across different arrays, which makes a direct comparison of these raw intensities difficult. However, by gathering statistics on these intensities and their spatial variation, a variety of metrics can be determined. In this study, the clutter distribution is found to fit better to a two-parameter Weibull distribution than Gaussian or log-normal distributions. Based on the spatial variation of the two Weibull parameters, scale and shape, more information may be gleaned from these data. How well the GPR array is illuminating various parts of the ground, in depth and cross track, may be determined from the spatial variation of the Weibull scale parameter, which may in turn be used to estimate an effective attenuation coefficient in the soil. The transition in depth from clutter- to noise-limited conditions (which is one possible definition of GPR penetration depth) can be estimated from the spatial variation of the Weibull shape parameter. Finally, the underlying clutter distributions also provide an opportunity to standardize image intensities to determine when a statistically significant deviation from background (clutter) has occurred, which is convenient for buried threat detection algorithm development that needs to be robust across multiple different arrays.

Detection of deliberately buried objects in ground-penetrating radar (GPR) data acquired along a path is a clutter-limited problem. Detection-false alarm rate performance can be improved by replacing the detection statistic with a change statistic that incorporates information from previous path traversals. A constellation matching approach is developed for buried-object change detection in GPR data. Network topologies of buried objects detected in GPR data from previous path traversals are maintained in a constellation database. Localized groups of buried objects newly detected on the latest path traversal are matched to the constellation. Buried objects from the latest path traversal whose locations or strengths cannot be reconciled with the constellation are identified as changes. The system has one component that generates constellation databases offline and another component suitable for change detection in real time. It can tolerate paths with significant translational misalignments. The system uses the following: 1) a customized translational relaxation algorithm for point pattern matching that incorporates detection strength and a probabilistic uncertainty model for buried-object location into the objective function and 2) a change statistic that accounts for the magnitude of change relative to predicted detection strength. A constellation database can typically be generated offline from a single path traversal roughly two orders of magnitude faster than the time typically required for a vehicle to travel the extent of the path. Database sizes are typically four to five orders of magnitude smaller than the data sets of GPR signal scans or focused 3-D GPR images that they were generated from. On bumpy dirt roads buried exclusively with nonmetallic objects at various depths, detection-false alarm rate performance is shown to be significantly better for our change statistics than for our detection statistics.

We extend the particle filtering approach to detecting polygonal structures of fixed shape and variable size in overhead images to the more practically useful case of unconstrained side lengths. Our approach deals with multiple candidate sides by validating and clustering particles based on evidence that corners of proper acuteness and orientation (as constrained by the polygon model) might exist. A queue-regulated tracking algorithm that handles multiple candidates across multiple sides is discussed. Compelling detection results in overhead images that involve entire families of polygonal structure are provided.

The problem of matching random tree models of multi-component patterns to tables or graphs containing components extracted from diverse data sources is considered. We focus on bi-level trees whose branches emanate from one root node and terminate on different leaf nodes. Node and branch attributes are treated as random variables. Tree nodes represent pattern components of specified types that occur in tables or graphs to be searched. For each item in the table or graph with a type match to the tree root, there is a set of components from the table or graph that are candidate leaves for optimal matches to the tree model. We adopt a view of optimal matches to random tree models as minimum cost assignments of candidate leaves to tree branches. Model-based formulas are derived for computing costs associated with assignments of specific candidate leaf components from tables or graphs to specific tree branches. We specify an ontology suitable for dynamic geo-spatial query problems in which (1) tree nodes represent physical objects or events on the ground (buildings, roads, communication transmissions...), and (2) branch attributes characterize, with uncertainty, distance or time separations between components, and angles between links connecting components. Our approach is used to search very large images for specific types of buildings in probabilistically constrained spatial arrangements, with the goal of ranking model matches for efficient inspection by human analysts.

In segmentation of remotely sensed images, the number of pixel classes and their spectral representations are often unknown a priori. Even with prior knowledge, pixels with spectral components from multiple classes lead to classification errors and undesired small region artifacts. Coarseness regulation for segmented images is proposed as an efficient novel technique for handling these problems. Beginning with an over-segmented image, perceptually similar connected regions are iteratively merged using a method reminiscent of region growing, except the primitives are regions, not pixels. Interactive coarseness regulation is achieved by specifying the area alpha of the largest region eligible for merging. A region with area less than alpha is merged with the most spectrally similar connected region, unless the regions are perceived as spectrally dissimilar. In convergent coarseness regulation, which requires no user interaction, alpha is specified as the total number of pixels in the image, and the coarseness regulation output converges to a steady-state segmentation that remains unchanged as alpha is further increased. By applying convergent coarseness regulation to AVIRIS, IKONOS and DigitalGlobe images, and quantitatively comparing computer-generated segmentations to segmentations generated manually by a human analyst, it was found that the quality of the input segmentations was consistently and dramatically improved

The problem of robust automatic road detection in remotely sensed images is complicated by the fact that the sensor, spatial resolution, acquisition conditions, road width, road orientation and road material composition can all vary. A novel technique for detecting road pixels in multisource remotely sensed images based on the phase (i.e., orientation or directional) information in edge pixels is described. A very dense map of edges extracted from the image is separated into channels, each containing edge pixels whose phases lie within a different range of orientations. The edge map associated with each channel is de-cluttered. A map of road pixels is formed by re-combining the de-cluttered channels into a composite edge image which is itself then separately de-cluttered. Road detection results are provided for DigitalGlobe and TeiraServerUSA images. Road representations suitable for various applications are then discussed

Human analysts are often unable to meet time constraints on analysis and interpretation of large volumes of remotely sensed imagery. To address this problem, the Image Content Engine (ICE) system currently under development is organized into an off-line component for automated extraction of image features followed by user-interactive components for content detection and content-based query processing. The extracted features are vectors that represent attributes of three entities, namely image tiles, image regions and shapes, or suspected matches to models of objects. ICE allows users to interactively specify decision thresholds so that the content (consisting of entities whose features satisfy decision criteria) can be detected. ICE presents detected content to users as a prioritized series of thumbnail images. Users can either accept the detection results or specify a new set of decision thresholds. Once accepted, ICE stores the detected content in database tables and semantic graphs. Users can interactively query the tables and graphs for locations at which prescribed relationships between detected content exist. New queries can be submitted repeatedly until a satisfactory series of prioritized thumbnail image cues is produced. Examples are provided to demonstrate how ICE can be used to assist users in quickly finding prescribed collections of entities (both natural and man-made) in a set of large USGS aerial photos retrieved from TerraserverUSA

The linear Parameter Plane Transform (PPT) of a height field attributes an inverted cone of empty space to each height field cell. In is known that height field ray-tracing efficiency can be improved by traversing rays in steps across inverted cones of empty space. However, steps across inverted cones of empty space along rays close to the base of a steep ridge will be short, even if there are no obstructions along the line of sight, because the cones will be narrow. This weakness can be virtually eliminated by allowing the opening angles of the inverted cones of empty space to vary between sectors, i.e., by directionalizing the linear PPT. An efficient algorithm for computing the linear directional PPT of a height field is given and its properties are investigated.

A new search method over (x,y,/spl theta/), called position-orientation masking is introduced. It is applied to vertices that are allowed to be separated into different bands of acuteness. Position-orientation masking yields exactly one /spl theta/ value for each (x,y) that it considers to be the location of a possible occurrence of an object. Detailed matching of edge segments is performed at only these candidate (x,y,/spl theta/) to determine if objects actually do occur there. Template matching is accelerated dramatically since the candidates comprise only a small fraction of all (x,y,/spl theta/). Position-orientation masking eliminates the need for exhaustive search when deriving the candidate (x,y,/spl theta/). Search is guided by correlations between template vertices and distance transforms of image vertices. When a poor correlation is encountered at a particular position and orientation, nearby positions at that orientation and nearby orientations at that position are masked out. Position and orientation traversal are by quadrant and binary decomposition.>

Height distributional distance transform (HDDT) methods are introduced as a new class of methods for height field ray tracing. HDDT methods utilize results of height field preprocessing. The preprocessing involves computing a height field transform representing an array of cone-like volumes of empty space above the height field surface that are as wide as possible. There is one cone-like volume balanced on its apex centered above each height field cell. Various height field transforms of this type are developed. Each is based on distance transforms of height field horizontal cross-sections. HDDT methods trace rays through empty cone-like volumes instead of through successive height field cells. The performance of HDDT methods is evaluated experimentally against existing height field ray tracing methods.

There are several methods available for representing boundaries in two dimensions. The more widely known ones involve boundary modeling with run-length codes, chain codes and polygons, AR model coefficients, Fourier descriptors (FD's) and B-spline control points (CP's). In this paper, we present results showing the usefulness of CP's in boundary representation. CP's can yield large compression of the boundary data and they exercise local control over boundary shape. Because they can be normalized with respect to scale, translation and rotation, they are valuable in boundary matching applications. Finally, being the series coefficients of spline basis functions, CP's are useful features for boundaries detected in noisy environments.