When the detectives on “CSI” can’t make out a license plate number in a grainy surveillance video, they bring it to the computer guy, who types a few lines of code and voila – the numbers are clear as day.

If only it were so easy. In reality, researchers have been laboring for decades to find ways of getting high-quality evidence from low-quality videos, according to Michael Black, professor of computer science at Brown University and an expert on computer-vision technology.

“It’s been a long road since the mid-’70s or so,” he said. “People thought it was going to be an easy problem to solve, and it’s turned out to be a remarkably hard problem.”

Although scientists have long been able to enhance simple images taken in controlled environments, it is far more difficult to do so when the source material is badly lit, real-world surveillance footage. Now, however, they are starting to crack the code of advanced forensic video analysis – and Brown researchers, led by Black, are at the forefront.

Following a workshop he gave on film restoration in 2006, Black was approached by a Virginia police department, which sought his help in solving a recent murder. Their best evidence – two surveillance videos – was so blurry that they could not even tell if the suspect’s car had two doors or four.

“They were desperate,” Black recalled. “They had asked everybody. No one could help them.”

Black threw out the curriculum for his Topics in Computer Vision class and instead had the course’s 16 graduate and undergraduate students spend the whole semester working on the Virginia video. By the end of the 13-week class, they were able to tell the police the car was a Toyota Camry made between 1992 and 1994. (Despite their efforts, the case remains unsolved.)

“It’s changed the way I teach entirely,” said Black, who now uses only real-world problems for the course. “You cannot believe how motivated students are to be solving a real problem. … They worked harder and were much more excited, and everybody was so committed to it.”

The key to video forensics analysis is comparing multiple frames of a video for clues, said Alex Weiss, 29, who works as a staff researcher in the computer science department.

“If you imagine a letter in a license plate and a camera on a telephone pole blowing in the wind, [the position] where the edge of that letter is will move back and forth across a couple of pixels,” he said. “If we can figure out in each frame where the edge should be, we can use the information in all those frames” to figure out what letter it is.

Black eventually approached R.I. State Police Lt. Dennis Pincince, who leads the Criminal Investigations Unit, about working together. They received a collaborative research grant from the R.I. Science and Technology Advisory Council, which paid for Weiss and another student to work full time on advanced video forensics at state police headquarters in Scituate last summer.

Pincince said the collaboration was particularly valuable because law enforcement officials have been using surveillance video more since the Sept. 11 attacks. Although image-enhancement products are available commercially, the price is often prohibitive for police departments. Plus, he said, “the stuff they’re working on at Brown is by far the best stuff I’ve seen out there.”

Black added: “It’s one thing to buy a piece of equipment – it’s another thing to have an expert [on the software there] who can modify it and fix it.”

Black and his fellow researchers are also developing a method for determining the body shape of an individual who is photographed wearing street clothes – even the baggy getaway garb of a petty thief.

The body-shape research has other possible uses as well – retailers could use it to demonstrate how clothing would look on an individual; gamers could create a realistic personal avatar and doctors could model athletes’ bodies to help in preventing injury. •