Imaging technology is no longer just about the never-ending megapixel race among CMOS image sensors. As market focus shifts to "vision" processing, the industry has drawn a new battle line.
CogniVue, Mobileye, CEVA, and Tensilica (now a part of Cadence) are just a few examples of IP companies enabling embedded vision technologies. The newest member to join the fray is Imagination Technology, which announced its PowerVR Raptor ISP (image signal processing) architecture Monday.
Leading chip companies such as Freescale, Texas Instruments, and STMicroelectronics are also rolling out purpose-built vision processors -- often taking advantage of their partnership/licensing deals with embedded vision IP vendors.
For the time being, though, automotive is the primary market for all these vision processors, since embedded vision is playing a key role in Advanced Driver Assist System (ADAS). Carmakers are banking on ADAS, advocating safety features such as lane departure warnings, collision mitigation, self-parking, and blind-spot notification.
According to IHS, a market research firm, revenue in 2013 for special-purpose computer vision processors used in under-the-hood automotive applications is forecast to reach $151 million, up from $137 million last year and from $126 million in 2011.
Hard problems to solve
I should, however, note that the industry is still scratching only the surface of the embedded vision future.
"Vision processing still remains as a very hard problem to solve,"Jeff Bier, founder of the Embedded Vision Alliance, once told EE Times, "despite the number of man-years spent developing a host of embedded vision algorithms."
CogniVue’s Wilson agreed. Processing a huge amount of real-time data demands intense compute power. To do a 3D sensor map in a robust manner, especially in a low-power consumer device, is especially tough, he added.
Asked why a 3D sensor map, he described it as "essential" to solve fundamental limitations in 2D computer vision. He noted that 2D, for example, has problems with segmentation (separating foreground from background), illumination (for face recognition), relative position (placing objects in the scene), and occlusion (hands in front of the face). Noting that different approaches for 3D sensing are fraught with tradeoffs, Wilson said that CogniVue is currently working on an algorithmic way to efficiently compute disparity maps for low-cost 3D sensor vision.
Designing hardware that can efficiently run different vision algorithms is a huge challenge for system designers. Options for system vendors looking for imaging/video processing solutions range from keeping it all in the CPU to offloading imaging to the GPU, or adding hardwired logic dedicated to imaging functions.
With the world's GPU IP core leader Imagination entering the vision market, the race among IP vendors and chip suppliers has only gotten even more intense.
There is no question that it's going to be a "Brave New World."