Get email delivery of the Cadence blog featured here
As I mentioned in my last post, I attended one day of the Embedded Vision Summit this week, and as I did last year, I had a great time learning about vision applications in the AI landscape.
Last year, I attended more of the “Fundamentals” and “Business Insights” track, which were more theoretical presentations, talking about what was possible with AI, machinelearningdeeplearning, and sensor applications. This year, I attended more “Technical Insights” presentations, which was basically about what has already been done with this emerging technology. We went from what is possible (last year) to what exists right now.
I attended “Computer Vision Hardware Acceleration for Driver Assistance,” by Markus Tremmel of Bosch, in which he explored the different hardware required for digital signal processing (DSP) of vision applications in ADAS systems. From using a distributed architecture (as is used in highway assist) to a centralized architecture (as will be used in true autonomous driving), there is a ton of data to be processed, and the camera offers the highest amount of data out of all the sensors used.
Click for more details. As you can see, others found these slides interesting, too!
From there, I attended Felix Heide of Algolux’s presentation, “Understanding Real-World Imaging for ADAS and Autonomous Vision Systems.” In it, he introduced a new standard, the IEEE SA 2020, the standard for automotive system image quality. He pointed out that real-world perception is difficult, what with everything from extreme lighting conditions, distortion from moiré patterns, weather obstruction, flickering LED lights, geometric distortions from wide-angle lenses, shadow—and the question becomes: what is “good enough” for use in an ADAS system? That’s what the IEEE SA 2020 attempts to codify.
The seven subgroups that go into IEEE-SA P2020
The next presentation, “The Roomba 980: Computer Vision Meets Consumer Robotics,” by Mario Munich of iRobot, was one I got excited about, and not just because I desperately want a Roomba of my very own. Years ago, I had an old version of the Roomba, in which the only sensors on it (that I could figure out, anyway) were the bumper on the front, the battery sensor to tell when it needed to be charged, and a signal that connected the base to the device so it knew where to go when the charge was low (but it frequently got lost on the way home). The theory was that it would take a seemingly random trajectory in a room, go in a straight line until it ran into something or it went too far, then stop, turn around in another direction, and go until it ran into something else; this pattern repeated enough was supposedly enough to clean the floor of an entire room.
A generic robot vacuum cleaner
The newer ones have so much more. (I am not a shill for iRobot, I’m just reporting what I saw!) The newer ones have odometry, gyroscopes, accelerators, cameras, bumpers, sensors to indicate what kind of floor is being cleaned, and others that I couldn’t write down fast enough. This little robot maps out the entire floorplan of the house it is cleaning, including clutter that may be in its way, and then cleans the floor systematically in straight lines, moving around table and sofa legs and beanbag chairs and your son’s discarded backpack and the cat the cat’s food bowl. When it needs a charge, it goes back to home base, charges up again, and then continues where it left off, until the “mission is completed”—the floor is clean.
Danger, Will Robinson, danger. Cat ahead.
The challenge in creating such a neat device is cost. Generally, when selling this kind of electronic device, it must cost only 20-30% of the retail price to build. So, for a $350 Roomba, it can’t cost more than about $60 to build. Mario, the presenter, went into detail about ways they were able to bring the cost down, including…
FOV = field of vision
In cutting corners wherever they can, they have created a Roomba that rivals the cost of any high-end vacuum cleaner.
My favorite slide of the day I wasn’t quick enough to take a snapshot of—Mario Munich was talking about how the usual design cycle involves R&D developing a product, and then throwing the design over the wall to production, where they take it and produce thousands of copies. The slide showed exactly that—throwing a design over the image of a brick wall. His point was, however, that production and R&D have to work together to make a product that works with such razor-thin profit margins.
I still want one.
The final presentation that I’ll write about is “Deep Understanding of Shopper Behaviors and Interactions Using Computer Vision", by Emmanuele Frontoni and Rocco Pietrini from Università Politecnica delle Marche. In the paper that they presented, they analyzed the behavior of shoppers around the world using little cameras embedded in the ceilings of several kinds of stores. The vision aspect of this research involves two different interactions: the macro, meaning how does an individual shopper interact with the entire store as they go through it, aisle by aisle; and the micro, meaning how does an individual interact with a specific item on the shelf. The latter metric involves three kinds of interactions: touching a product but not taking it, taking the product, or putting the product back on the shelf.
Using the ceiling cameras, they were able to track the behavior of thousands of people, and come up with meaningful data to give to, among others, Proctor & Gamble, about how stores should be laid out, what items should be stored at eye level, and what visual marketing seems to work for specific classes of product.
We should all be aware that our very gestures are being tracked and recorded; I think the illusion of privacy is long gone. My son was just telling me the other day that he would like to have an intelligent app that tracks his day for him—how much time is he spending staring off into space, for example, or watching television, or messing around on his phone, or talking with friends, or on personal grooming or petting the cat or whatever. He said it would really help him to budget his time. But my first thought? It seems like it would be a huge invasion of privacy. What if someone other than he got hold of that information? I’m sure marketers would have a field day with that kind of data.
But when I pointed that out to him, he kind of shrugged. “They pretty much have all that information, anyway,” he said.
I suppose it could be a generational thing. But I’d rather track my own life, thankyouverymuch. Still, if there were an app for that on your phone, vision sensors and machinelearningdeeplearning will play a huge role in it.