New System Enhances Images In Crime Investigation
Software originally developed by MacDonald, Dettwiler & Associates (MDA) to help Mars rovers get around obstacles has been refined into a prototype system that can convert a few seconds of video from a handheld stereo camera into a 3D model of a location. The software is being tested by Canadian and U.S. police departments as a tool for aiding crime investigations.
MDA’s Instant Scene Modeler software improves on earlier computer vision systems because it can function in natural light, process data faster, and operate without the need for special markers in the environment to indicate where frames should be stitched together, thus eliminating concerns about disturbing the scene.
The software builds the scene in 3D using the local invariant features technique, in which the most distinctive portions or features of a photo are extracted, and information about their location, orientation, size, and brightness is encoded in a small file.
Descriptions of several hundred features for each photo are stored in a database that a computer can search when it is faced with an unfamiliar image; the software can accurately choose a matching image even if as much as 90 percent of the features do not match.
The software splices images together using common features in each frame as indicators of where to join frames, and new video can be spliced into the model after it is constructed. MDA’s software employs University of British Columbia professor David Lowe’s Scale Invariant Feature Transform algorithm, which Evolution Robotics has also licensed for use in its computer-vision software.
Microsoft researchers are developing software that can turn digital photos into a panorama using their own invariant features algorithm, while similar algorithms for identifying objects or sites and assembling 3D models from video and photos are also under development.
Abstracted by the National Law Enforcement and Corrections Technology Center(NLECTC) from the New York Times (03/10/05) P. E8; Ricadela, Aaron .