Much attention has been paid to large-scale AI models, whose usefulness in perimeter security remains to be seen due to certain challenges and limitations.
AI video analytics can play a key role in perimeter security. Especially, AI can be effective in threat detection and behavioral analysis, which are important in keeping the perimeter safe and secure. Meanwhile, much attention has been paid to large-scale AI models, whose usefulness in perimeter security remains to be seen due to certain challenges and limitations.
Perimeter security refers to the protection of an organization’s perimeter or outer boundary. For many organizations, perimeter security is critical as it serves as the first line of defense. In this regard, AI-based video analytics can be quite useful, especially in the following areas:
Intrusion detection/object analysis
AI can do a good job performing intrusion detection and object analysis, which are key in perimeter security. “For example, AI-driven intrusion detection helps operators identify attempted breaches of perimeter infrastructure and access control, and with AI analytics software, they can identify both subtle anomalies and active threats. Object detection offers further insight into potential threats, warning operators of aggressive weapons associated with criminal activity,” said Todd Dunning, Director of Product Management, Video Security and Analytics at
Pelco.
Behavioral analysis
AI can also be effective detecting abnormal behaviors that may suggest a potential breach against the perimeter. “Cameras informed by AI video analytics software can not only identify actions consistent with break-ins, but also signs of loitering and suspicious movement that could suggest a planned perimeter breach in the future,” Dunning said.
Vehicle gate misuse
Not just individuals, vehicles can also do damage to the perimeter, and detecting and deterring ill-intentioned vehicles at the outside is important. “[AI can be helpful in] identifying vehicle tailgating, forced entry, or wrong-way approaches, along with license plate recognition for automating entry and exit for authorized vehicles,” said Mats Thulin, Director of AI and Analytics Solutions at Axis Communications.
The aforementioned use cases can be enabled with various AI technologies including computer vision, license plate recognition, facial recognition, objection recognition/behavioral analysis technologies and others.
“Today’s core technologies include advanced enhanced imaging – such as denoising, HDR, low-light optimization, and image stabilization – along with AI-driven object detection and classification for reliable detection of people, vehicles, and animals. Audio analytics can also play a key role by identifying sounds like glass breaking or screams,” Thulin said. “Together, these technologies provide operators with a fuller, more comprehensive and dependable view of the perimeter.”
Large-scale AI models
Recently in security, much enthusiasm has been generated in the topic of large-scale AI models, which are said to be quite effective in perimeter security. For example, it is said that large-scale AI models can detect targets in longer distances, recognize objects in much smaller pixels and identify a wider range of animal species. While large-scale AI models are quite promising as a next-generation security technology, their effectiveness in perimeter security is still debatable, given certain challenges and difficulties.
“Large models require intense cloud processing, introducing 15-20 second delays minimum. In security scenarios where attackers move in milliseconds, this delay is operationally useless,” said Albert Stepanyan, President and CEO of Scylla AI.
He adds: “Large AI models are architected for single-frame analysis, not continuous video streams. They miss critical temporal patterns – a person climbing a fence over 3-4 seconds appears as disconnected static images. This fundamental design flaw creates massive false negative rates where genuine threats go undetected.”
According to Dunning, both large-scale AI models and the more conventional convolutional neural networks (CNNs) have merits and demerits, and adoption depends on the use case.
“CNNs provide a more reliable and resource-efficient solution for typical installations where operators know which types of threats they face and are confident in their abilities to adjust to meet unique, pre-specified needs. Large-scale models are best-suited for high-risk, well-funded sites that often face unknown risks and have access to vast resources,” he said.
Thulin of Axis, meanwhile, said a hybrid approach may be the way to go.
“Large-scale, general-purpose models may impress in demos, but perimeter security demands real-time performance in outdoor, edge environments where speed, predictability, and power efficiency are critical. Today, compact, purpose-trained models tuned for specific tasks and environments are the most effective and lead to less false alarms. Also, many of the larger models are difficult to run on edge devices requiring server or cloud processing and in the latter case this also creates a dependency on cloud connectivity,” he said. “Long term, a hybrid approach is ideal: leveraging larger models in the cloud for second level processing, while deploying efficient analytics at the edge for live detection in the perimeter use case.”