Nick Bowden, Managing Director of Digifort UK, a video management software (VMS) and analytics developer, looks at AI’s impact on the CCTV industry now and in the future.

AI development is core to video analytics performance in CCTV, improving effectiveness, accuracy, channel density and reducing cost.

However, the HPC (high performance computing) hardware development which enables the processing escalation that AI demands, will also affect physical CCTV system capability, design, layout and cost.

AI now

Software

In CCTV applications, AI and analytics are synonymous.

The video analytics technology used in CCTV solutions can broadly be split into three categories; Neural, Deep Learning and Binary.

Digifort supports all three types, as well as integrating with the analytics in third-party NVR’s, analytics boxes, and cameras at the ‘edge’.

Each has a cost performance consideration, so Digifort works with them all to ensure customers have flexibility, for example, not spending on advanced analytics where cheaper motion detection might do.

Leading analytics types use Machine Learning (ML) to train the software algorithm to recognise and interpret objects in a scene, including relevant movement and behaviour patterns.

Like human recognition, many different objects can be identified, with a ‘confidence’ figure, from a stored library of known objects, learnt by the system over time.

These objects might include people, vans, bikes, cars, trucks, groups of people, bags, cyclists and many more, including their colour profiles and movement directions.

State of the art, Deep Learning (DL), a category of ML, increases accuracy and effectiveness further.

This includes the ability to automatically self-calibrate and discount scene items that are of no interest to the analytics, reducing false alarms.

Another example is overlaying skeletal frames on people to track the position and movement of hands, arms and heads relative to the torso, along with speed of movement, to identify more complex behaviour patterns such as aggression or violence.

Hardware

Current VMS applications with analytics use servers with CPU (Central Processing Units) for the VMS operation, with enough grunt to process video from the system’s cameras.

Servers or PCs with GPU (Graphics Processing Units or Graphics cards) provide the computer processing capability needed for analytics.

The Digifort VMS will easily process 100 to 200 cameras on a single, 2U server with a mid-range, Intel Xeon CPU-type processor, depending on the camera recording profiles.

Digifort, an Nvidia partner, designs its analytics to run on Nvidia GPU, where a mid-range graphics card, such as the RTX A2000, will process up to 60ch of VA, depending on the type.

So, both the VMS and VA processing are done using established, well-proven hardware to deliver cost-effective, CCTV system building blocks.

Network, storage and the cloud

Perhaps the greatest limitation in CCTV applications currently, in both physical hardware terms and understanding, is network bandwidth and storage; critical to AI and analytics.

If a 4MP camera on a remote site is streamed at 25FPS, using an efficient H265 compression algorithm, its bandwidth might be around 3Mbps, just as an example, (depending on scene activity, image quality and camera type).

When this video is recorded for 31 days, it will need around 1.0TB of storage.

An 8ch CCTV system will, for example, need a multiple of that; 24Mbps and 8TB.

I am not sure what the average broadband connection delivers these days, but 100Mbps down speed and 20Mbps up speed is far better than my current residential connection, and that would not be enough, as it is the up speed we need when streaming from a remote site to a central location.

Digifort offers rental license options and supports centralisation on a remote server or cloud (someone else’s server), but the cost of cloud storage and the broadband connection necessary to stream effectively are expensive.

Intelligent VMS systems like Digifort have variable bit rate, adjusting camera bit rates in real-time to deliver the bandwidth where it needed most.

However, unless the mindset moves from continuous recording to event recording, triggered by AI-driven analytics in real-time, for larger systems, this dog won’t hunt, as they say! And that’s before we factor in latency.

AI in future

Software

In future, AI will help video analytics improve performance and predictive analysis, such as accuracy, speed, breadth of objects recognised, behaviour patterns and logic of CCTV, as well as efficient integration with third-party systems for improvements to site management.

This might include asset protection and monitoring, access control, intruder, business and building intelligence, PSIM and more, but also IOT mass data processing to predict potential security threats through analysing patterns and trends in behavior to highlighting exceptions.

In addition, the quantity of channels per site will grow, enabled by computing processing speed and power.

Industrial revolution is used in the title of this article, because AI has reached a critical point, well demonstrated by Chat GPT.

This large language AI model, developed by OpenAI, will write text and find images, trawling the interweb for relevant information and presenting it in a readable, logical format, within seconds.

The same type of AI model is being used for software development, where the AI can help optimize itself, such as to improve VA algorithms, without human intervention.

Likewise with VMS processing and AI hardware design, the human bottleneck has been removed and AI is improving itself!

Hardware – processing

The current generation of AI hardware is electronic based.

However, AI is creating an explosion in demand for greater bandwidth, low latency and faster processing capability.

This requires High Performance Computing (HPC), currently in development for datacenter applications, but no doubt migrating into wider computing applications soon.

Optics-based ‘photonics’ connections are appearing, as Co- Packaged Optics (CPO) technology seeks to replace copper connections with optics – vastly increasing processing speed, reducing system size, power consumption and operational costs.

Nvidia joined the exclusive, Trillion-dollar company club this year on the back of developments for AI technology.

The CCTV system ‘building block’ sizes mentioned earlier of 100 ~ 200 VMS channel VMS and 40 ~ 60 VA channels, naturally align with larger CCTV systems and can often be cost-effectively deployed on site.

However, if HPC technology enables next-generation VMS and VA processors to handle just 10x the channel count, within the next 5 years, processing and possibly storage must be centralized, to a remote server of datacentre (cloud), with cameras allocated to a site and paid for on a per site or per camera basis, to be viable.

Next-generation datacenters are expecting a massive uptake in processing demands for many industries, coming to the same realisation, as HPC drives a remote processing model.

Clearly NVRs; boxed analytics solutions; and analytics cameras simply can’t keep pace.

Network, storage and cloud

To effectively centralise, AI-based analytics need the network to have zero latency to avoid delays between streamed video and analytics events triggering.

5G, optical networks should provide zero latency and are in the short-term pipeline.

If integrators and end users can accept event driven CCTV to reduce bandwidth and data storage costs, remote solutions for storage and analytics processing will prevail.

If recording must be continuous, then hybrid solutions with local storage and centralised VA processing, may result, with cost reductions in local storage delaying the move to cloud storage and streaming for larger systems.

Summary

The development of AI is driven by consumer products, smartphones, gaming and many further markets, with CCTV benefiting as a result.

Developments in hardware, with huge computer capability, will make AI-based analytics perform better.

However, large channel counts mean processing exceeds that of typical systems, so must be spread over multiple sites.

A centralised or cloud CCTV system allows costs to be spread, but network latency and cloud storage costs, along with broadband streaming costs may keep storage local for longer.

Either way, an upskilling of CCTV engineers, with greater optical and network capability, will be required.

This article was originally published in the August Edition of Security Journal UK. To read your FREE digital edition, click here.