Some New Tricks

AI-based cameras have some surprises up their sleeve

• By Adam Lowenstein
• Mar 19, 2024

In the constantly evolving world of AI, you can blink and miss an innovation. This rapid pace of evolution means organizations are under increased pressure to invest in solutions that do not become obsolete a short time after they are installed.

For video security in particular, AI evolutions in machine and deep learning are consistently bringing new products and services to help us do a better job of protecting people and assets. Like our smartphones, features and functions can evolve rapidly when we consider AI apps running on edge devices such as a camera. An app-based ecosystem allows for flexible customization, updates and deployment of features tailored exactly to the job at hand.

For this reason, it has never been more important for companies to seek out vendors and manufacturers with open platforms that collaborate freely with multiple third-party manufacturers.

AI has fundamentally changed video security. We have gone from cameras that can sense basic motion, which were prone to false positives from passing shadows or wind-blown trees, to highly accurate human and vehicle detection with descriptive attributes and real-time alerts in just a couple of years. Further evolution has led to cameras that can detect anomalies in a scene (scene change detection) such as a door being left open, or a vehicle left in a no-parking spot beyond a preset time limit. The same analytics can even notify when stock is running low on shelves.

Modern AI cameras have become flexible IoT devices. And like our smartphones, we can now think of them as platforms for hosting specific applications for the unique job required. What will the product designers and engineers think up next?

How about AI cameras that can be trained to recognize custom objects on site? It is even possible for an AI camera to analyze the video of non-AI network cameras, effectively turning them into “new” AI cameras as well.

Customizable AI On-site Learning
While AI-based security cameras have been able to significantly reduce errors by reliably detecting humans and vehicles for some time, this next phase of AI is irresistible to data-hungry businesses. Customizable AI on-site learning enables integrators and end-users to train a camera’s AI analytics on-site to recognize unique objects that are important for a business to track or count—precisely what so many customers have asked for.

On-site AI training can further enhance accuracy by recognizing logos on vehicles or uniforms, counting planes, forklifts, baby strollers or shopping carts. This new stream of business intelligence data, harvested directly using edge processing within security cameras, enables more automated workflows while increasing operational efficiency and enhancing service quality.

For example, operators could teach the camera to count forklifts or shopping carts passing through the camera’s field of view to provide new metrics about operational efficiency. A hospital can count ambulances arriving at the emergency room. The camera can even be taught to recognize a logo on a truck and send out an alert when it arrives at the loading dock. The best AI on-site learning apps can even auto-generate multiple training images at different luminance values saving operators valuable training time while further increasing detection accuracy.

By processing custom data locally on the edge, AI on-site training removes the need to send sensitive data to the cloud for analysis. This is particularly beneficial for industries handling confidential information, such as healthcare or finance, as it reduces the risk of data breaches and ensures compliance with data privacy regulations like GDPR or HIPAA.

Extending a Camera’s Life
It is no secret that AI helps security operators effectively monitor hundreds if not thousands of cameras for events of interest. This is so valuable, that it is easy to imagine a not-too-distant future where every camera includes AI capabilities. Today, there are millions of network cameras installed, and most do not have any AI features. While it is possible to send all those streams to servers for external AI processing, it can be much cheaper, faster and more secure to process some of those non-AI streams using existing AI cameras on the edge.

The most recent AI cameras are powerful enough that they can not only analyze their own video streams but can also analyze and extract valuable AI attributes from traditional cameras that lack AI capabilities. This innovative capability allows customers to add AI features to their existing, non-AI, surveillance cameras, including cameras from different manufacturers, making them smarter, more efficient, and able to trigger real-time alerts.

This cost-effective solution improves existing surveillance systems and further reduces false alarms in a phased approach without requiring forklift upgrades. Network cameras which previously had limited AI integration to popular VMSs like Milestone, Genetec and Video Insight can now pass AI metadata exactly as if they were the latest AI model. This is exceptionally powerful for cameras that may be installed in difficult places to retrofit.

The most powerful AI cameras can currently process up to three additional video streams from non-AI cameras. Doing the math, that is one new AI camera turning three traditional network cameras into AI cameras as well.

A Collection of Apps
In the same way our smartphones run a unique and curated collection of apps, modern AI cameras can be considered as platforms that host custom AI applications for the unique tasks required for any organization. Customizable AI on-site learning enables integrators and end-users to train a camera’s AI analytics on-site to recognize unique objects that are important for a business to track or count such as forklifts, shopping carts, or even airplanes.

The most recent AI cameras are so powerful, that they not only analyze and extract AI metadata from their own video streams but can also analyze and enhance traditional cameras with the same AI capabilities. With the rapid pace of AI evolution, make sure that any future investments in physical security are not obsolete mere days after they are installed. Look for open platforms that promote flexibility and customization — a solution that plays well with others.

This article originally appeared in the March / April 2024 issue of Security Today.