Artificial Intelligence Embedded in a Camera

Pushing AI to the Edge

Horizon Robotics claims to be a leading technology powerhouse of embedded Artificial Intelligence.  The company provides integrated and open embedded Artificial Intelligence solutions of high performance, low power, and low cost.  Their website states “We equip smart cameras and cars with “brains”, turning them into intelligent entities that have the ability from perception, understanding to decision-making for convenience, safety and fun.”

Brilliance Security Magazine sat down with Horizon Robotics Founder and CEO, Dr. Kai Yu to learn more about how embedded AI can be used in surveillance.  To be fair, we didn’t test their cameras or even validate the claims made.  There are others in the industry much better suited for that kind of product analysis than we are.  Our intent here is to offer a vision of what is technologically possible by coupling the latest developments in microprocessors and AI.

In simple terms, AI is a machine that perceives its environment and takes actions that maximize its chance of successfully achieving its goals.  One tool used to create AI is the artificial neural network.  Inspired by human or animal biological neural networks, these systems “learn” without task-specific programming.  In computer image recognition these commonly called “neural networks” might learn to identify images that contain cats by analyzing sample images that have been manually labeled as “cat” or “no cat” and then using the results to identify cats in other images. They do this without any prior knowledge about cats, e.g., that they have fur, tails, whiskers and cat-like faces. Instead, they evolve their own set of relevant characteristics from the learning material that they process.

Horizon Robotics offers a camera with sufficient processing power to run its own neural network.  In a very real sense, the entire camera now becomes an artificial intelligence.  It is a device capable of learning and making decisions that will enable it to attain its specified objectives.  With the ability to process 1080p video at 30fps it can detect, identify, and track up to 200 objects – faces maybe – per frame.  With a typical power consumption of only 1.5W, you now have a powerful server built into each camera.  A device able to provide facial recognition or any other analytic you desire, at scale; think finding the bad guy in a stadium full of potential faces.

The advantages that immediately come to mind are the economics of bandwidth and processing power.  If the camera is an AI device it will send back to the headend, or to the cloud, only the images or frames that meet the objective – only what the operator is looking for.  And, if the processing is done at the edge, no robust server is needed to process and analyze the video stream.

Much of Horizon Robotics work is in autonomous vehicle development in China.  As noted by Bloomberg “The [Chinese] government’s aspiration to deploy 30 million autonomous vehicles within a decade is seeding a fledgling chip industry, with startups like Horizon Robotics Inc. emerging to build the brains behind those wheels.”  There once was a time when cars and cameras seemed so far removed from each other that a company operating in both spaces would not make sense.  Today, however, autonomous vehicles depend on AI and smart cameras to make nearly every decision.

It seems likely that processors running neural networks embedded into cameras to create intelligent surveillance devices will be the way of the future.  The possibilities are astounding.  However overused the buzz-word “AI” is today, it offers the potential to completely change how we use and design surveillance systems in the future.

Steven Bowcut, CPP, PSP is the Editor-in-Chief of Brilliance Security Magazine