Neural network approach could revolutionise video analytics

Ambarella is a big player in the video surveillance market, but not a familiar name to many buyers of security cameras. They don’t make cameras, but they make the computer chips inside. Founded in 2004, Ambarella began in the broadcast infrastructure encoders market and entered the market for professional security cameras in 2008. More recently, the company has also entered the market for automotive OEM solutions.

Between 2005 and 2015, the company has produced a progression of advanced camera system on chips (SoCs) designed, developed and mass-produced for the consumer electronics, broadcast and IP camera markets. An SoC includes an image processor as well as capabilities to run software and provide computer vision (analytics).

Development has been happening fast at Ambarella. In January, they introduced the CV22 camera SoC, combining image processing, 4K and 60fps video encoding and computer vision (video analytics) processing in a single, low-power-design chip. CVflow architecture provides DNN (deep neural network) processing required for the next generation of intelligent cameras.

The even newer CV2 camera SoC, introduced in late-March, delivers up to 20 times the deep neural network performance of Ambarella's first generation CV1 chip, also with low power consumption.

I caught up with Chris Day, Ambarella’s vice president of marketing and business development, at the ISC West show to find out more about the company.

Q: Your company is not as well known in the industry as it should be, given its widespread impact on the market. Would you prefer otherwise?

Day: I think we would prefer more visibility. If you talk to any camera maker, they know who we are. We do business with all the top-10 camera companies – Hikvision, Dahua, Avigilon, Pelco and the rest. Because we are a chip supplier, the end-customer deciding to buy a camera may not know what chip is inside. For that reason, we may not have the visibility. But if you are a camera maker, you know who we are.

Q: What are you hearing from your camera customers in terms of what they need, and how are they directing where you go with R&D?

Day: We have become a major supplier to those companies based on years of developing image processing – wide dynamic range, low light, and similar features – as well as AVC (advanced) and HEVC (high-efficiency) video encoding. That’s the heritage of our company and why we do business with all these companies. The next treadmill is computer vision – adding the intelligence into the camera. The goal is still being best-in-class in imaging and encoding, but now being best-in-class in adding the intelligence and being able to do all those things with very low power, within the “thermal budget” of the camera. That’s the next big wave.

Q: How far away is that in terms of the end-customer? How soon will he or she be able to reap the benefits?

Day: By the end of 2018, or maybe next year. We’re just beginning to sample the CV22, for instance, which is the first SoC directed to security cameras. Typically, it takes nine months to develop a camera, maybe longer with an intelligent camera because you are importing so much software. So, we’re talking about the end of this year or next year.

Q: Tell me about your current products and the next generation.

Day: The CV22 is sampling this quarter. CV2 we announced [in late March], which is a high-performance chip. The idea is that we provide our customers with different price/performance points, so they can produce a family of cameras with different capabilities. They have the same basic software model, so someone can invest in software once and then have different performance points without completely rewriting the software. That’s key. They might have 100 software engineers developing neural networks and all the features, so if you have to recreate that at different price points, it’s a lot of work.