Pattern Recognition in
For more information on topics in this article:
Professor Michalopoulos’ web page:
Photos of the cameras on Highway 407 in Toronto:
Toronto 407 ETR web site:
Web site of Declan McCullagh, photographer:
Technical papers describing Autoscope®:
Image Sensing Systems’ web page:
While driving north of Toronto last winter, I took a turn onto Highway 407, a toll road. As I drove, I was surprised to see no tollbooths, and I did not to have to stop or slow down to pay a fee. Instead, there were frame structures over the road with cameras mounted on them. As a pattern recognition person, I was pleased and impressed at the technology that evidently captured license plates by which to allocate toll charges. A few months later, while driving in London, I took a wrong turn into the “congestion zone” in London, for which drivers require a permit – which I did not have. I reversed my path as quickly as I could, but not before a camera had captured my license plate. I found this out later as I received a ticket in the mail. I was less pleased with the technology this time, but still impressed by its efficiency and determined to find out the story behind it.
Dr. Panos Michalopoulos is a professor of civil engineering at the University of Minnesota whose field of expertise is traffic engineering. He is also founder of Image Sensing Systems, a pioneering company in the development of machine vision systems for traffic management and control. I spoke with him to learn more about these cameras, their early development, and their use of machine vision and pattern recognition techniques.
As Panos described to me, automobile detection devices are not new. The first sensor dates back to 1928, a simple switch that made contact due to the car’s weight as it drove overtop. The devices used commonly today are magnetic or inductive loops, placed just beneath the road’s surface, that detect the metal of a car. You can observe these by the rectangular cuts in pavement, often in lanes entering an intersection. Loop detectors have been used since the 1950s. As effective as this technology has proven to be over the years, there are several shortcomings. Once a loop is inserted in the road, it cannot be moved. To obtain extra 2-D spatial and 3-D space-time information, more sensors must be installed: an expensive solution.
In the early 1980s, Panos was already very familiar with the problem of traffic engineering. His Ph.D. thesis and early research dealt with models of the problems and issues involved in this area. Although theoretically very efficient, there was an element lacking: there was no way to prove and use the solutions because they required more information than the inductive loops could practically provide. Panos knew that he needed a better means for capturing real traffic scenes, but his civil engineering education provided no background to the sensing solution. Through a colleague at Honeywell, he learned of the field of machine vision. It was evident that the solution lay within this area, but the field was in its adolescence. At that time, machine vision was barely able to track a single object in motion. Furthermore, it was constrained to small or low-resolution images due to lack of computing power. Undeterred, Panos and his colleagues founded the company Image Sensing Systems and undertook the challenge of building practical vision systems for traffic detection.
Those readers who have designed image analysis systems will recognize familiar challenges. The first challenge was to separate cars from the road background. Easier said than done as they struggled to make the systems robust with respect to day and night lighting, rain and snow backgrounds, and shadows and reflections. Another challenge was the absence of fast, inexpensive, and small computers. To meet time and cost requirements, they developed an innovative solution, which they called detection lines. Cars were not recognized and tracked in 2-D; instead they were detected as they crossed these lines. This solution reduced computation substantially because the 2-D analysis was reduced to several 1-D analyses, as several detection lines could be placed in a single image to capture traffic flow. Furthermore, unlike physically static loops, these lines could be placed and moved interactively by traffic engineers at a computer console (versus a paving crew on location!). By 1989, they had a prototype traffic detection box, which consisted of one or more cameras, a processor board, and weather-resistant housing. The device was called an Autoscope.
Panos describes the first major installation in Oakland County, Michigan, as having inauspicious beginnings. There was a giant American flag at an intersection where the Autoscopes were installed. The flapping of the flag frequently obscured the camera views causing utter confusion for the analysis algorithms. This noise had not been anticipated. There were many more learning experiences as theory and laboratory design were adapted for real conditions. By the early 1990s the system was robust and began selling well.
With much more powerful computers today, 2-D zones, rather than lines, are placed in the images. Via multiple virtual detectors in an image and multiple cameras, 100 or more sub-images can be captured and processed by a single processor. Personnel can place and change locations from a remote location. The detection signals can also be enabled and disabled by time or by captured events. More sophisticated algorithms can measure speed, volume, road occupancy, and vehicle classification. This information can be important not just for traffic control, but for several other traffic-related tasks such as monitoring for accidents and congestion, collecting traffic data, and reducing air pollution and fuel consumption.
Oakland County now has more than 800 intersections outfitted with Autoscopes. Throughout the world, there are over 35,000 cameras deployed in more than 50 countries.
Now, when you drive onto a booth-less toll road (or enter London’s congestion zone), be assured that your car will be detected. You can thank the pioneering work of researchers like Panos Michalopoulos who have applied machine vision and pattern recognition techniques to this area of traffic engineering.
Feature Articles from the IAPR Newsletter
Pattern Recognition in Origami, Jan. ‘05
Pattern Recognition in Origami, Jan. ‘05
ICPR2004 Invited Talks, Oct. ‘04
Pattern Recognition in Maps, Sep. 03
Pattern Recognition in Sports, Apr. ‘03
What is Pattern Recognition?, Winter 2003
All photos in this article are from the Toronto Highway 407 ETR (Express Toll Route).