| |
    
|
A new computer system makes roads safer for both cars and pedestrians
by Bill Goin
Most of us have been stuck at a traffic light that is red for no reason-no traffic, no pedestrians, just the mechanical workings of a clock. We've also gone through a green light only to find a pedestrian who doesn't believe that traffic signals apply, forcing us to swerve or slam on the brakes. And we have also seen elderly or disabled folks trying to cross the street only to have the light change before they have the opportunity to finish crossing.
Professor Nikolaos Papanikolopoulos, of the University of Minnesota's computer science department, has a plan to change all of that. He has developed a system that would link traffic signals to real world conditions. Instead of a timer that is oblivious to the environment, there would be a computer that would detect pedestrians and react accordingly.
Because computers and people process information differently, this task, which would be trivial for a human operator, requires a sophisticated computer system. Identifying pedestrians requires a large amount of processing power.
Papanikolopoulos's system involves a camera, a powerful processor, and a sophisticated computer program that is able to convert raw images into symbols that represent pedestrians and, in some cases, vehicles.
Because its signal is easily converted to data that can be processed by a computer, a Charged Coupled Device (CCD) camera is used. A CCD camera can produce 30 frames per second to be analyzed by the computer system. A lot of processing power is required because each image contains "millions of bytes" of information, according to Papanikolopoulos.
The system operates with three levels of data. First, the raw data is transmitted into the system. Then, the computer subtracts the background and views only images that change. Finally, the computer analyzes the non-constant data and determines whether it is a pedestrian (which the system is concerned with) or something unimportant.
A busy intersection has background, streets, buildings, parked cars, traffic lights, fire hydrants, and the like. In front of that are things that move -- people, pets, cars, trash, etc. To facilitate the determination of which images are associated with pedestrians, the computer first subtracts a "static" background image, which greatly reduces the amount of information to be processed when identifying pedestrians. The system has the "ground image" (background) stored in its memory. It subtracts this ground image from the raw data (or "current image") to produce the "figure image," which involves changing images.
Papanikolopoulos explained, "The ground image changes over time to adapt to light, time of day, weather, and other things." The system is sophisticated enough to create a new background so that the processor is only concerned with moving images. Depending on how rapidly the environment is changing, the background image can be replaced every few hours or, if necessary, every few minutes.
While the current image from the CCD camera is a high-quality black and white image with various shades of gray, the figure image is in black and white. Placed on a monitor, it looks like white blobs on a black background. Trying to determine which blobs are pedestrians and which blobs are not requires a sophisticated program. "The characteristics of an image have to correspond to one of the models of a pedestrian in the system's memory," Papanikolopoulos said, adding that "pedestrians have over twenty different ways of moving."
The computer analyzes "thousands of images in a time sequence. It uses the information over time and sees if it matches the movement pattern of a pedestrian," explained Papa-nikolopoulos. In the computer, pedestrians are represented as active, deformable models. They have a stick shape with two legs, two arms, and a head. Papanikolopoulos noted, "The system focuses on special characteristics of a pedestrian. Specific points, like feet, are good to focus on and make good landmarks."
In Papanikolopoulos's lab, the viewer can observe all three images--current, ground, and figure--simultaneously. When the program is running, the computer's "pedestrians" are shown as rectangular symbols on the current image. This allows you to see how well the system's model corresponds to reality.
In the early part of the project, pedestrian models needed to be entered into the computer database. These models were developed from people moving around. Osama Masoud, a graduate student assisting Papanikolopoulos, explained, "We needed human guinea pigs to portray a variety of movements in order to help create the signatures of pedestrians."
After the system identifies an image as a pedestrian, it follows the image. Sometimes there is an obstacle between the camera and part of the field of view, which creates a blind spot. Masoud said, "In blind spots the system assumes the pedestrian continues at the same course and speed." As a result, the system occasionally creates a pedestrian where there isn't one, but it can adapt and symbolize a new pedestrian when someone who stopped, or changed direction, behind an obstacle re-emerges into view. Masoud stated that the system is adaptable. "Sometimes two or more people will enter the field close together and be perceived as one pedestrian. If they separate, the computer can split the image into several pedestrians."
One of the proposed uses for this system is to control traffic lights at pedestrian images. Masoud explained that "the light will stay green only until all pedestrians have crossed or until a certain maximum time has elapsed." He also noted that it will start the warning signal in time to allow the slowest pedestrian in the intersection time to cross the street. Papanikolopoulos added, "If there is someone in the intersection, oncoming traffic won't see a green light, but maybe a flashing green to warn them."
Although the technology is sophisticated enough for a field test, it hasn't been used yet. "There are legal issues," Papanikolopoulos said. "Before you deploy an instrument like this you have to decide who is responsible in case of an accident." To date, no one has agreed to accept this responsibility.
