Canadian system enhances image: C.S. investigation

[Tami]

Canadian system enhances images in crime investigation
By AARON RICADELA

New York Times News Service — Forensic experts who reconstruct crime scenes want to produce detailed drawings that can stand up in court without disrupting sensitive evidence.

But creating hand-drawn sketches and taking photographs can take days and disturb the scene.

Computer-aided design software that require investigators in the field to enter data can be cumbersome, and results can be difficult for jurors to decipher.

Now, a Canadian company is demonstrating prototype software, based on advances in computer vision, that can stitch together a few seconds of video from a hand-held stereo camera into a detailed 3-D model of a room, including the people and the objects in it. Using Windows on a laptop, the police or courtroom workers can zoom around the model to view it from different perspectives, or click on its features to see sizes, relative distances, areas and angles.

"It gives you a third dimension," said Detective Inspector Jeff Wilkinson of the Ontario Provincial Police's criminal investigation branch. He is assessing the prototype for MacDonald, Dettwiler & Associates. "You can actually begin to develop your investigative theories about how something transpired without entering the scene."

Based in Richmond, British Columbia, the company is best known for designing the robotic arms used on the space shuttles and the International Space Station. It originally developed its Instant Scene Modeler software to help Mars rover missions negotiate obstacles, then adapted it for new markets.

The company, which has been testing the system with police departments in Canada and the United States, hopes to release products within a year that can also be sold to mining companies for mapping excavations, museums for constructing virtual tours and the military for guiding unmanned vehicles, said Frank Teti, product development manager.

The software is based on a new technique, called local invariant features, for comparing computer images. It quickens processing times, works in natural light and does not require special markers to tell the computer where to stitch together frames.

Researchers with other companies are using the technology to recognize objects instantly in a shopping cart, guide robotic toys and construct panoramic photos from overlapping images.

Earlier computer vision systems worked by trying to compare all the pixels in two images or by matching small regions of images called templates. Image comparison runs into trouble if it encounters motion or changes in lighting; template matching suffers if the templates get rotated, are lighted differently or change in resolution. Template matching works well in controlled environments, but it is slow and uses lots of processing power.

Instead of comparing pixels, the invariant features approach pulls out the most distinctive patches of a photo, called features, and encodes information about their location, orientation, brightness and size in a small file. A database holds descriptions of several hundred features for each photo, which a computer can search when it encounters an unfamiliar image. The software can accurately select a matching image even if only 10 per cent or so of the features match.

"The evidence accumulates fast," said Paolo Pirjanian, chief scientist and general manager for robotics and vision at Evolution Robotics in Pasadena, Calif., which uses the invariant features approach in its computer-vision software. That speeds processing time and compensates for changes in rotation, size and brightness.

To build a 3-D scene, the MacDonald, Dettwiler & Associates system uses common features in each frame as anchors to show the software where to join images. The more video shot, the better the model. Users can also splice new video into the model after it is built.

Technology like MacDonald's is enabling new applications that were impossible just a few years ago, said David Lowe, a computer science professor at the University of British Columbia who developed an algorithm called SIFT, or Scale Invariant Feature Transform, that MacDonald, Dettwiler & Associates and Evolution Robotics have licensed.

"Images are becoming so common with digital cameras, cellphone cameras and medical scanners," he said. "We just need ways for computers to automatically interpret all that information."

Sony's latest Aibo robotic dog uses Evolution software based on SIFT to find its way back to its charger and take commands from coded cue cards. An Evolution anti-theft product for grocery stores called LaneHawk can identify items on the bottom shelf of a shopping cart from a database of about 5,000 photos within a quarter of a second, without using a bar code scanner or requiring customers to lift them up.

Pietro Perona, an electrical engineering professor at the California Institute of Technology who also sits on Evolution's scientific advisory board, is researching software that can categorize similar photos after it recognizes them. That could help search engines group pictures by their content, he said. "Recognition of objects was extremely crummy before SIFT features came out, and now it's very reliable," Perona said. "You're able to think about products, whereas previously you couldn't."

Researchers at Microsoft have developed a local invariant features algorithm, which allows Microsoft's Digital Image Suite 10 and Digital Image Pro 10 software to create a panorama from standard digital photos, and they plan to discuss the technology's details at a computer vision conference in June. Microsoft is also working on similar algorithms to recognize objects or locations, and build 3-D models from still photos and video, said Richard Szeliski, a senior researcher and manager of Microsoft Research's interactive visual media group.

[url=\"http://www.globetechnology.com/servlet/story/RTGAM.20050310.gtcrimemar10/BNStory/Technology/\"]Source: Globe Technology[/url]