Thanks to the latest AI-powered video surveillance systems, modern security cameras can do far more than just record and store footage. As part of smart surveillance solutions, they actively analyze their environment, identify unusual activity and respond to potential threats in real time.
Unlike traditional video analytics, AI-powered security systems—often built into modern video management software (VMS)—serve as semi-autonomous partners for human security teams. With features like real-time object detection, facial recognition and behavioral analysis, the advent of AI marks a great leap forward in the accuracy, speed and reliability of security operations.
Aside from what the tools can do, the origins of AI models play an increasingly large role in the decisionmaking of users seeking to upgrade their systems—because of cybersecurity and compliance calculations, as well as their accuracy in, for example, detecting faces.
A question of money
Ever larger shares of the R&D budgets of security cam brands are therefore dedicated to AI research. In many cases this means the bigger the company, the bigger their AI research budget, the better their AI models.
Developing AI models at the edge of what’s currently possible in-house is obviously costly. Smaller manufacturers, such as those that are focused on entry-level and mid-range cameras and equipment, might struggle to keep up.
They basically have three to four options if they want to stay in the AI race:
- They can buy camera modules from OEM partners that are preloded with third-party AI. The modules are often “white label,” which means companies can brand them as their own, without crediting the actual manufacturer where the user sees it. Brands usually have little oversight about the AI that’s in them.
- They can license AI software development kits from others, including dedicated AI developers, but also security industry giants such as China’s Hikvision or Dahua, and customize the AI. Once again, they will only have limited in-depth knowledge of what’s at the core of “their” AI.
- Or they can use open-source AI and customize it. This, too, requires fewer resources and know-how, but it has the advantage that open-source models enjoy greater trust. It’s unlikely that an AI of this kind has unknown backdoors that can be exploited.
The provenance of any AI, regardless of whether it’s in a device from a billion-dollar brand or a boutique manufacturer, is usually difficult to trace, and companies refrain from fully opening up about it. The deeper you dig, the closer you get to trade secrets that companies justifiable want to keep to themselves.
AVTECH's in-house pivot
Taiwan-based AVTECH choose Option 4—it developed the main parts of its AI in-house, even though it's costly.
“We try to do it mostly ourselves, but, for sure, we cannot do it 100%,” Alejandro Portillo, Product Division Sales Manager at AVTECH, told asmag.com during an interview at Secutech.
For the rest, they rely on trusted partners in the public sector, such as Taiwan’s Industrial Technology Research Institute (ITRI) or National Chiao Tung University (NCTU), which both have a proven track records in AI research.
Thanks to this close relationship with local R&D partners, “we can say proudly that we offer fully Taiwanese solutions,” Portillo said, adding that this is especially important in markets like the US or India, where Chinese security technology is largely banned.
“The only issue is that [customers] sometimes expect Taiwanese products, but with Chinese prices,” Portillo said.
Local vs cloud storage
AVTECH's main product lines, which it showcased at Secutech from May 7-9 in Taipei, include recorders that can handle medium-sized security infrastructures with up to 64 cameras. It allows users to process and store data locally without having to rely on subscription-based cloud services. Many AVTECH customers would find monthly subscriptions too costly, Portillo said.
AVTECH also allows users to integrate legacy devices, such as older analogue cameras, which, as Portillo explained, also helps users keep costs down.
On the other hand, AVTECH's self-developed mobile application software EagleEyes features a host of AI-powered tools. These include facial detection and recognition, human detection, vehicle detection and automatic license plate recognition.
Relatively straight-forward AI tools can go a long way, said Portillo.
“It is not only for security purposes, but you can also [use AI to get] business insights—for example, to count the flow of customers inside your retail business,” he explained. “It’s a trend for us to find business applications and develop functions that go beyond security.”
At the core of the AI
Common security AI tools—offered by AVTECH and many other companies, such as facial recognition—are based on deep learning models (typically convolutional neural networks) trained to classify visual inputs.
Basically, it’s like this: The camera “sees” something, and the AI decides whether or not it’s the face of a person who is already in the database—a relatively simple “yes/no” operation paired with a binary response based on large amounts of “AI training data.”
These datasets might come from different sources—third-party developers, in-house or public/open source. AVTECH uses a mix of the latter two.
Other AI-powered features, such as AVTECH's fall and fire detection tools, which automatically alert the user if an incident is detected, are based on slightly more complex operations. They require temporal analysis—i.e., understanding movement or changes over a span of time, not just in a single frame.
Such tools are based on spatiotemporal deep learning models, or combinations of motion tracking, object classification and—once again—the application of binary rules.
AI tools that make “yes/no” decisions on the basis of relatively straight-forward deep learning models are indeed what’s behind most AI features found in security systems—in the mid-tier equipment of smaller manufacturers as well as high-end gadgets of premium brands.
Vanguard AI
AI tools that are at the edge of what’s currently possible, such as “context-aware” AI or AI-powered communication, are mostly available in higher-end equipment. They usually come at a larger price, and require more computing power, which also raises the costs, especially if they're run locally—that means without relying on cloud infrastructure.
Context-aware AI is, for example, at the core of smart city architectures that can interpret multiple layers of data (time, location, behavior, the identity of subjects, historical data) to make nuanced decisions. It's not just about detecting an object, but understanding what is happening in a broader context to identify threats before they become apparent to human security teams.
Meanwhile, AI-powered communication in security devices mostly comes as user interfaces that resemble chat bots. Users can, for example, punch in “show me all old men on bikes” and the AI delivers results without requiring security personnel to select predefined categories.
AI on top of AI
AVTECH offers a more traditional, Linux-based user interface—one that its longstanding partners are used to, Portillo said.
The company's devices do, however, allow for AI-powered natural communication options—if they are paired with third-party VMS that has the respective AI built in.
“Developers of smart search software are among our customers as well. They have the software, we have the hardware,” Portillo said, adding that even cloud-only companies offering the latest AI eventually rely on hardware partners.
Any cloud-only AI that runs on AVTECH devices is not covered by the company's “Made in Taiwan” pledge, though. It obviously depends on what AI other components from other manufacturers use.
It once again holds true: The more intricate the AI infrastructure, the more difficult it is to know where it all comes from.