Since the inception of networked video surveillance, many companies have worked to develop a variety of different analytics to enhance the value of the systems to end-users. Some vendors have been more successful than others in being able to provide reliable video analytics to their customers and, after a period in which the technology was greeted with a healthy amount of skepticism, it has now become commonplace in many surveillance installations across a wide range of vertical markets.
While the use cases for analytics have changed, the technology itself has remained relatively the same – algorithms are created to search for certain pre-defined actions within a camera’s field-of-view. However, the evolution of artificial intelligence (AI) means that the future of analytics will lie not in the creation of static algorithms but on the ability of machines to learn what operators should and should not be alerted to. For example, for those installations that use virtual trip wires for notification of perimeter breaches, many analytics cannot decipher between a human coming onto the property, which would obviously be the primary concern, vs. an animal, which would be of little interest. The market is not oblivious to this, however, as both industry stalwarts and start-ups alike are investing in analytics based on machine learning technology.
Last week, Movidius, a provider of embedded machine vision technology, announced that its Myriad 2 Vision Processing Unit (VPU) will be integrated into a new lineup of smart cameras from Hikvision. The Myriad 2 will be used to run “Deep Neural Networks,” which enable surveillance systems to automatically detect anomalies at the edge inside the cameras themselves rather than at the server level or in the cloud.
“Traditionally, the bulk of video analytics have been run in the network, whether it was in the NVR or even further down the stack. The problem with this approach is that you end up with a single point of failure and you’re at a real disadvantage because the real-time aspect is particularly impacted when everything is cloud-based,” Movidius CEO Remi El-Ouazzane says. “Our technology provides the ability to deploy these workloads locally on the device to maximize performance in a way that makes sense for IP network cameras that would not be possible otherwise.”
El-Ouazzane says those who deploy their technology can expect to achieve greater accuracy with analytics they currently leverage. In the future, he says organizations will be able to implement surveillance systems that predict crimes before they occur. “You are going to have the ability to have much better accuracy for known workloads, which are things like vehicle identification, plate recognition or the ability to see if someone is wearing a safety belt or not. These analytics are used today but are not as accurate as people want them to be,” El-Ouazzane explains. “A neural network will allow you to improve accuracy from anywhere between 25 to 75 percent.
“More importantly, you are now going to have the ability to modelize behavior. Neural networks and supervised learning allow you to create a system that is able to potentially prevent things from happening,” he continues. “For example, if you have enough data sets in a retail store that you can use to actually create a behavior of someone on the verge of committing a theft, you can use different neural networks against this and essentially have the ability to move in a world where you, as a store manager, are being made aware of a theft before it has even happened because you were able to predict it based on behavior modelization.”
In addition to Hikvision, Dahua and Uniview have also inked deals with Movidius to use the company’s technology inside some of their cameras. It’s not just larger industry firms, however, that are looking to advance video analytics through the use of AI. Earlier this month, tech start-up TeraDeep, a developer of deep learning applications and AI appliances, announced that it has created a new product that will offer a dramatic improvement over the performance of traditional video analytics. Unlike conventional solutions that use graphical processing units (GPUs), TeraDeep leverages a field-programmable gate array (FPGA)-based architecture that offers faster analytics at half the power.
The first version of the company’s solution is an FPGA-based PCIe board, developed in conjunction with tech firms Xilinx and Micron, which achieves a four-time lower latency compared with the latest GPUs. The board also runs TD Accel, TeraDeep’s deep learning acceleration technology. The advantage of using TeraDeep’s FPGA architecture running on neural networks means that the FPGAs can take on the computing-intensive tasks of rapid general categorization (identifying whether something is human or animal) and then pass off the more complex task of differentiation (who the person is) to a GPU, which results in significant compute performance and power savings.
“Our product is a complete solution,” TeraDeep CEO Didier Lacroix says. “We combine hardware in the form of FPGA cards with our own deep neural network technology that we’ve developed for a vast range of applications. On top of that, we also have an API server where people can tap into the AI/deep learning capabilities by writing a little bit of code.”
The company is currently targeting the government and defense markets with its products, but they also have plans to develop similar solutions that could be deployed within the commercial security sector on a more wide-scale basis in the future. In fact, Lacroix says the company has already done some work with one prominent provider of body-worn cameras along these lines, however, he believes the mass market simply isn’t ready for the type of performance offered by TeraDeep as of yet.
“We are ten times faster than GPUs in terms of latency,” he says. “Our unique value proposition is really built around the ability to detect very fast moving objects.”
Moving forward, Lacroix believes that artificial intelligence is going to be a critical component of video analytics across every vertical; however, he says it is also important that companies not overpromise and under deliver as has been the case in the past.
“From a government/defense side, this is part of the requirements now. They understand the benefits and they know they have to start to integrate this technology as part of the solutions that are being deployed,” he says. “On the commercial side… I think it is something that needs to be approached carefully.”
