Author(s): Jeremy Green
The potential for visual analytics is enormous. The gathering of image and video data through Internet-connected cameras, and the intelligent processing of that data either at the network edge or in the cloud has the ability to complement, or perhaps replace, sensor data in many cases. From face and object recognition in public surveillance systems, through monitoring in assisted living solutions, smart parking occupancy detection, to retail footfall analysis... the list really is endless.
The diversity of these scenarios is breathtaking in terms of the characteristics of the data and the requirements of the application. In the smart parking example, it's sufficient to determine whether there's an object of the correct size in the specified location; if there's a big rectangular thing there, then the parking space is occupied. But picking out and recognising a specific face in a crowd of moving people is much more demanding; quality control on a factory production line might be somewhere in the middle. Sometimes the camera needs to move about, sometimes it must be able to pan, sometimes it can stay fixed.
If it weren't for the fact that we humans derive much of our primary sense data about the physical world from two paired light-sensing organs, we probably wouldn't think about all of these different applications as belonging to the same "vision" category at all.
This diversity means that creating solutions is best handled by an open-ended, flexible platform with interfaces to other components and to the widest possible spread of developers, rather than by a series of closed vertical systems. So, Qualcomm's announcement on 11 April of its Vision Intelligence Platform seems to fit the bill rather well.
The platform includes two new system-on-chips, the QCS605 and QCS603, which feature an on-board image sensor processor, Qualcomm's Artificial Intelligence Engine, an ARM-based multicore CPU, vector processor and GPU. The platform also encompasses Qualcomm's camera processing software, machine learning and computer vision software development kits, and connectivity and security technologies. The announcement builds on Qualcomm's launch of a reference design for a Snapdragon-based IP camera in October 2016, which itself incorporated on-board processing capability for machine learning and artificial intelligence uses.
And although it's a vision intelligence platform, there are high-end audio capabilities, including noise and echo cancellation, on-device audio analytics and processing features for natural language processing, audio speech recognition, and "barge-in" capability, which supports voice interfaces in noisy environments.
There are already some customers preparing to develop products based on the platform, including Ricoh and IP video specialist Kedacom. Perhaps more importantly, there's a raft of technology provider partners, including Pilot.ai, MM Solutions, and SenseTime, which has $1 billion of backing from Alibaba and was described last week by Bloomberg as "the most valuable artificial intelligence start-up in the world".
The new platform is in several ways a bold move by Qualcomm. The QCS605 and QCS603 support full Wi-Fi and Bluetooth connections. As we've indicated before, the Internet of things domain has multiple and diverse requirements for connectivity, so this approach is to be welcomed. And it builds on the company's rather distinctive platform approach to the Internet of things, of combining connectivity and computing power into segment-oriented packages.
It's bold too, in that others have a much stronger and better established position in the image- and video-processing space; Hikvision, the video surveillance specialist company ultimately owned by the Chinese government, comes to mind as the most obvious contender. So does Movidius, acquired by Intel in 2016, which similarly offers powerful devices capable of processing at the edge, and which has some impressive agreements in place with Google and Microsoft, among others.
Qualcomm will face a tough climb in establishing itself in this domain, and will find itself in a not entirely familiar position of being a challenger rather than a market leader. There's no shame in that, but abandoning the entire sphere of vision analytics to others would be something of a surrender.
A version of this article first appeared in Techworld on 30 April 2018.
Sign up to our free Daily Insight service here.