Learning to see

Automotive Design & Production,  April, 2004  

• E-mail
• Print
• Link

Vision system developer Mobileye N. V. (Amstelveen, Netherlands; www.Mobileye.com) has a pretty simple business plan: become the Intel of the automotive industry. The way it proposes achieving that modest goal is through the near-universal adoption of its new EyeQ vision system-on-a-chip as the heart of the next generation of advanced driver warning systems. (For an overview of these systems see ADEP December 2002, "Danger Ahead" [http://www.autofieldguide.com/articles/120203.html].) Mobileye execs reckon that EyeQ puts the company three years ahead of its competition; plenty of time to secure the loyalty of the world's automakers and first-tier electronics suppliers and put "Mobileye Inside." So, what's so great about this system that it should cause such, ah, visions of success? First, EyeQ's system-on-a-chip design packs a lot of image processing power in a compact footprint, with the equivalent of two Pentium-level processors that have been optimized for pattern recognition and 300 KB of on-chip memory that can store several frames of video at a time. Second, it's comparatively inexpensive, using only one low-cost high-dynamic-range CMOS (complementary metal-oxide semiconductor) camera which mounts in the base of the rear view mirror. (While the company is not divulging the system's cost, chairman and chief scientist Amnon Shashua does say, "$100 is the magic number that allows a system to enter into high volume," implying that EyeQ either meets or beats that sum.) But the factor that Mobileye thinks really gives it an edge is the way its system processes images.

The software that drives EyeQ learns to recognize objects the same way the human brain does, by seeing enough examples to form classifications: road, tree, sign, moose, etc. It then learns to estimate distances and time-to-contact based on its experience. Of course, the method requires thousands of images to be fed into a computer to teach it not only what a highway looks like, but, for instance, what it looks like in heavy traffic at dusk. However, according to Shashua, this method, though labor-intensive at first, is far less complex and expensive than setups that require two cameras to be constantly calibrated with each other to achieve proper depth perception.

[ILLUSTRATION OMITTED]

More broadly, Shashua claims EyeQ has a leg up on the laser-based systems that have been popping up in adaptive cruise control units on high-end vehicles, because not only is it much cheaper, but it can decipher the richness of images in the real, crowded world. "Radar cannot classify objects," he explains, "A Coke can or a truck is the same thing to it." But EyeQ can not only tell the difference between soft drinks and semis, it can segment images precisely enough to tell if they are being viewed from the side or the front, which could prove crucial for applications like pedestrian detection.

Mobileye says most of the major automakers and Tier One electronics suppliers are on its client list, with the first production models equipped with its system due out in 2006. "Each automaker has a different angle on the technology," declares Shashua, "Some only want lane keeping functions, others want forward collision warnings or pedestrian detection, it depends on the maker and the market." Of course, the bottom line for Mobileye is that regardless of the specific application used, its chip has to drive it, bringing it that much closer to Intel status.--KEW

COPYRIGHT 2004 Gardner Publications, Inc.
COPYRIGHT 2004 Gale Group