This paper summarizes a system, and its component algorithms, for context-driven target vehicle detection in 3-D data that was developed under the Defense Advanced Research Projects Agency (DARPA) Exploitation of 3-D Data (E3D) Program. In order to determine the power of shape and geometry for the extraction of context objects and the detection of targets, our algorithm research and development concentrated on the geometric aspects of the problem and did not utilize intensity information. Processing begins with extraction of context information and initial target detection at reduced resolution, followed by a detailed, full-resolution analysis of candidate targets. Our reduced-resolution processing includes a probabilistic procedure for finding the ground that is effective even in rough terrain; a hierarchical, graph-based approach for the extraction of context objects and potential vehicle hide sites; and a target detection process that is driven by context-object and hide-site locations. Full-resolution processing includes statistical false alarm reduction and decoy mitigation. When results are available from previously collected data, we also perform object-level change detection, which affects the probabilities that objects are context objects or targets. Results are presented for both synthetic and collected LADAR data.
Structural segmentation of 3-D point-cloud data is an important step in the acquisition, recognition and visual representation of objects from point data. Associating groups of points that are consistent with structural surface elements allows decision making based on object components that are much more meaningful that the points alone. Processing begins by filtering the 3-D point-cloud data to smooth surfaces and remove noise. Filtering is essential for accurate surface-normal estimation. Our point filtering algorithm steps a 3-D box through the data, using an efficient search algorithm that employs priority queues for sequential sorting in x, y, and z. Filtering is based on the computation of a best planar fit at each box location. After filtering, processing continues by again
stepping through the data and computing local surface normals at each filtered point. We then compute a Minimum Spanning Tree (MST) with nodes consisting of the filtered points, edges established by proximity, and edge weights set as the Euclidean distance between local surface normals. A modified range tree that is
computed on the fly from unsorted point data is used in implementing the MST. We then employ a novel procedure to determine the edges that should be broken, leaving subgraphs that represent structural surfaces. These surfaces have been used for visual display of 3-D LADAR data, extraction of surfaces for automatic detection of buildings, and differentiation between man-made and natural objects.
We present a new technique for the automatic segmentation of multiband imagery. Our approach is based on the computation of a minimum spanning tree over a graph derived from the image. We use a fast graph-search algorithm and a custom tree-splitting algorithm to provide a high level of performance. This approach captures some of the gestalt characteristics of human perceptual grouping and is easily adaptable to a variety of spectral and spatial criteria. We demonstrate our technique on multispectral and hyperspectral imagery of the earth's surface.
A multispectral normalization processing system has been developed to produce percent reflectance maps from multispectral imagery (MSI) in the .4 to 2.5 micron wavelength range. It is adaptive to multiple spatial resolutions, supporting resolutions in the .25 meter to 30 meter range. The normalization process takes advantage of known naturally occurring and man-made materials in the image to remove the effects of atmospheric haze and sensor gain contributions for each multispectral band. The output product is a percent reflectance map for each multispectral band. Although the normalization technique is well known, the MSI normalization system (MSINS) provides a simple, adaptive, robust graphical user interface for normalizing multispectral imagery from various sensor platforms. Over 130 different surface material spectra have been collected from reputable sources in literature and other spectral material libraries and installed in the MSINS Materials Spectral Information Database (MSID). The MSID has been designed to allow the addition of new material spectra into the system via a menu interface. A neural-net-based region grower has been developed to minimize user interaction and increase the robustness and repeatability of the normalization. New multispectral sensor platforms can be introduced into the system quickly via a menu interface. The current system was developed and tested using Landsat Thematic Mapper, Erim M7 Mapper, Positive Systems ADAR 5500, and ITRES casi multispectral imagery.
Use of remotely sensed imagery to map lines of communication or revise existing maps is currently a labor-intensive process. In this paper, we present a system for automatically extracting lines of communication from high- resolution (less than 5-meter spatial resolution) multispectral imagery. Positive systems ADAR 5500 imagery is used to demonstrate system functionality. Our system includes automatic detection and identification algorithms, a geospatial database for storage and retrieval of results, a change detection component that compares newly detected lines of communication against stored database information, and a user interface that allows operator review and editing of automatically extracted results.
A new technique for the extraction of lines of communication (LOCs) from Landsat Thematic Mapper (TM) data is presented. A multi-stage approach is taken. First, LOC segments are detected. Next, gaps between segments are filled by a segment connection routine. Finally, the connected segments are identified. In the segment detection stage a distinction is made between wide LOC and narrow LOC segments. Wide LOC segments are detected by a neural network based segmentation algorithm. Input to the network are spatial and spectral features computed about each pixel. Narrow LOC segments are detected by a cost minimization algorithm designed to work on the multispectral data. Segment connection is performed by connecting those LOC segments that have consistent alignment, position, and spectral signatures. The identification stage consists of a neural network with inputs of shape and spectral features about each connected LOC. While the system is still being refined, most critical pieces have been prototyped and tested. Initial results are encouraging.