This article looks at some of the common analytics used in today’s AI cameras, as well as technological advances that make running intelligence on the edge possible.
AI cameras, as their name suggests, are cameras with AI-based analytics built into them. Powered by these analytics, AI cameras can help users gain better security and business intelligence. This article looks at some of the common analytics used in today’s AI cameras, as well as technological advances that make running intelligence on the edge possible.
AI cameras are seeing more popularity and market acceptance these days. From detecting suspicious objects to recognizing VIP customers at retail stores, AI cameras can be quite useful in various scenarios. Meanwhile, increased demand for AI cameras has triggered their growth: According to Grand View Research, the global AI camera market size was estimated at US$13.93 billion in 2024 and is projected to reach $47.02 billion by 2030, growing at a CAGR of 21.6 percent during the period.
Common analytics found in AI cameras
AI cameras are powered by various AI-based analytics that help users meet their security and non-security objectives. Some of the more common analytics found in today’s AI cameras are summarized as follows.
Object detection
Most modern security cameras come equipped with AI-based object detection. “This enables them to identify whether something is a person, vehicle, or other object. This capability is commonly used to trigger rules like tripwires or perimeter alerts and to generate metadata that can be used by the video management system (VMS) for tasks like forensic search,” said Florian Matusek, Director of AI Strategy and MD at Genetec Austria.
Behavioral analysis
Behavioral analysis is another common feature, allowing cameras to determine what someone may be doing and whether danger is present. “Today, the majority of edge cameras, including domes, bullets, turrets, and PTZs, can provide highly accurate AI-powered automated functions, such as motion detection, intrusion detection, line crossing detection, loitering detection, and fall detection,” said Sean Um, GM and Head of European Business Development Team at IDIS.
Face/license plate recognition
Facial recognition and license plate recognition can help users achieve certain security and non-security objectives such as recognizing VIP customers or allowing cars to enter frictionlessly into car parks. “While facial recognition and license plate recognition (LPR) are also available on some cameras, those tend to be used in more specialized, vertical-specific applications,” Matusek said.
Attribute extract
AI cameras also include advanced attribute recognition capabilities. “These include distinguishing the gender and age of persons of interest, as well as whether they are wearing glasses, hats, or masks, or carrying accessories like backpacks and bags. The technology can also distinguish between various vehicle types (such as sedans, trucks, buses, and motorcycles) and colors,” Um said.
What enables these analytics
The aforementioned analytics can be quite complex and may require extensive compute resources to run. Fortunately, more and more of today’s AI cameras feature powerful, advanced systems-on-chips (SoCs) that can handle these analytics
“Many modern IP cameras, particularly those from manufacturers like i-PRO, are built with powerful edge processors designed specifically for real-time AI workloads. They can handle multiple analytics streams concurrently and support containerized applications, allowing end-users to deploy custom or third-party AI tools tailored to their needs. As an example, a single i-PRO X-Series cameras can process a stream of AI analytics in the camera plus an additional three video streams from non-AI cameras for a total of four streams on one device,” said Adam Lowenstein, Americas Product Director at i-PRO.
“High-performance IP cameras are now being built using powerful SoCs with integrated neural processing units (NPUs) or even with separate NPU chips (dual-chip architecture). This enables edge devices to support more advanced AI features directly on the camera,” Um said. “Most mid-to-high-end IP cameras released in the last few years can run a range of analytics, thanks to improved processing hardware. For the most demanding AI tasks, however, external processing either in the cloud or on specialist NVRs and servers will still be necessary to be effective.”
A word on large-scale AI
It should be noted that certain vendors claim their cameras run the so-called large-scale AI models to “understand” instead of “see” scenes. The reality, however, is whether today’s AI cameras have enough compute resources to execute these models effectively is still questionable.
“The size of CNN models that a camera can run on-device depends on the performance of the NPU and DRAM capacity – larger models require a more capable NPU. Running transformers (such as GenAI, LLM, VLM) typically requires specialized transformer HW acceleration in the camera SoC’s NPU – particularly when there is a need for both high performance and high efficiency,” said Shay Kamin Braun, Director of AIoT Product Marketing at Ambarella.
“While today’s high-end cameras can handle fairly advanced deep learning models for tasks like object detection and classification, they aren’t designed to run large-scale or generative AI models, such as those used in natural language processing or complex scene understanding,” Matusek said. “The kind of processing power required for these more sophisticated models is still beyond what current camera hardware can support. The chips that could enable this level of performance on the edge are only just beginning to emerge. So, for now, if you want to run large-scale AI models, that still needs to happen on a server or in the cloud.”
Product Adopted:Others