printer friendly version

AI developments in surveillance

When AI-powered video analytics first emerged in the surveillance market, it was heralded as a game-changer, promising near-magical object recognition and identification. However, initial expectations were not fully met, and the performance of AI-powered video analytics often disappointed. Yet, the potential of AI in this field remains high, and with continued advancements, we have seen and expect more improvements in the future.

Despite the challenges that object recognition and identification pose for AI-powered video analytics, the potential of AI technology in the surveillance field is undeniable. Factors such as changes in lighting, complex background patterns, and occlusions can be overcome as AI technology evolves and improves. Today, some are even claiming that their AI can reliably recognise humans and other objects, even with cheaper or old analogue cameras supplying the images.

To find out what is really happening, SMART Security Solutions approached three companies focused on AI and asked for their input on what is happening in their world. We were not asking for the ‘magic’ the marketing people promise, but what AI does in the real world and how it really performs. The people respondents included:

Dr Jasper Horrell

Gerhard Furter

• Andre Vermeulen, Secutel Technologies.

• Gerhard Furter, Iris AI.

• Dr Jasper Horrell, DeepAlert.

What’s happening in the AI world?

We start by taking a broader look at what is happening in the world of AI in general and what companies are doing to take advantage of advances and market requirements in the security world. What can AI providers confidently offer the security industry today?

“The AI landscape has experienced a revolution in recent years,” Vermeulen says. “At our company, we have harnessed these advancements to deliver what was once considered impossible. Three years ago, we offered basic analytics. Today, we provide a comprehensive AI suite that unlocks a whole new level of security and efficiency, including features like human pose detection and a new human model. This evolution in AI technology is a testament to its potential and the exciting possibilities it holds for the future.”

“Over the past few years, the use of AI has become more generally accepted by customers and the general public,” adds Horrell. “Initially, as AI providers, we would ignore areas using exclusion zones. However, we have moved to alert on zones of interest, as opposed to ignoring areas as a whole.”

“We can confidently offer clients more accurate detection due to improved architecture, as well as larger datasets. The use of large language models means a wider variety of options are available compared to simple person and object detection.”

Furter believes there are three major evolutions that occurred in the AI world that directly impacted the industries today:

• The cost of computing hardware came down, thus impacting the cost of an AI deployment and making AI a viable alternative to ‘smart’ detection systems.

• The field of machine vision became a respected field of research, leading to a surge in new abilities and features.

• The market became educated, thus fostering a new sense of expectation that upped the standards for what is accepted as ‘AI’.

Clients do not just accept promises; they now know what to look for and demand noticeable results.

Is cheap really good?

As noted above, some promise that your cheap no-name-brand camera is good enough for excellent AI analytics, but what are the realistic expectations for cameras, or what quality image does an AI need to really do its job?

Based on his experience with Iris AI’s own products, Furters states that the following constitutes “a good AI camera spec”:

At least a 640 x 480 image for basic human and vehicle detection, but optimally 1280 x 720 will provide dependable results with almost no false positives. Light-enhanced cameras are better, otherwise, night-time image analysis may become a challenge. We can do object classification, ANPR, weapon detection, body posture analysis, person tracking and counting, and customised object detection on such an image, with acceptable success when facial recognition is applied. However, we found that a 1920 x 1080 image is perfect for excellent facial recognition.

• We also found that these deployments are markedly dependent on the client’s network and offsite communication.

Horrell explains that a realistic specification would need a camera with an adequate lens to send DeepAlert a high-quality VGA resolution event. “Although VGA is advisable, our model does cater for CIF resolution and lower, with higher degrees of sensitivity. When choosing a camera, bandwidth processing costs can also be high. Using very high resolution and sending thousands of images is going to cost more. It’s better to look at a mid-range camera and lean on the AI software to compensate a little.”

He continues that more sophisticated models will rely on GPU processing power, and the more GPU power, the more you can do with AI. What is essential is diversity in the training images in the datasets fed into the software. Effective machine learning and training can compensate enormously for cameras that may have a lower resolution or lighting issues at certain times of the day, etc. “Striking a good balance between the hardware and the software is key, so you have an effective setup for your specific scenario.”

Therefore, the answer is also, “It depends”. Vermeulen says, “Not all cameras are created equal for AI. While a top-of-the-line 4K model might seem ideal, our expertise lies in identifying the optimal camera setup for your specific needs. We consider factors like camera placement, lighting, and environmental angles to ensure your AI solution captures the critical details.”

What about the not-so-clever among us?

While AI services in every industry aim to make it as simple as possible for customers to use, this is not always possible. AI is complicated, and getting the process right is never easy; it is not simply installing and using an application. While the people we asked for input may be able to understand what is happening behind the scenes, the rest of us, mere mortals, rely on others to make it easy to use.

Most people would like a ‘set-up-and-go’ AI solution, perhaps with some fine-tuning in the initial stages – such as virtual ‘walk-throughs’ or something similar for specific locations and tasks. Can simplicity be provided as an out-of-the-box solution, or is training a necessity for each site?

Due to its large datasets, Horrell says that DeepAlert offers a ‘set-up-and-go’ AI solution. “We have trained most of our cameras in all potential scenarios, so training is not required to offer our filtering service. However, training may be required around new scenarios, new AI models we wish to offer and perhaps where there is a very different environment or model to be used. Customers support the training process by having the ability to flag events that are seen as nuisance events/false alerts.”

Similarly, Secutel offers pre-configured AI solutions for a quick start, however, Vermeulen notes, “We highly recommend training your AI for your unique location and tasks. This empowers your system to adapt to your specific needs and maximise its effectiveness. Many potential customers are unfamiliar with AI training, and that is where we excel. Our team will educate you on the benefits and guide you through the entire process.”

Iris AI has simple, self-install products, such as SnapGuard (https://snapguard.co.za/), which requires no technical knowledge beyond basic internet skills. However, ActivEye (http://www.activeye.co.za/) will require skilled technicians to install, according to Furter.

“None of our products require additional training, just parameter tweaking in the form of motion sensitivity, paranoia and who to send alarms to, etc. We have found, however, that customers are eager to train, and often even expand their training by implementing their own small ‘tricks’ to better their deployments.”

Looking ahead

As a constantly evolving technology, we also asked our respondents what their expectations of AI are over the next three years (or so). What will customers be able to rely on that they cannot right now? More importantly, will development be so fast that there will be new features and functionality available that are impossible now?

Iris AI believes that AI will become more commoditised and become a household technology. In addition, Furter says the level at which custom data sets will function will become easier, and it will become a simple process to implement specialised classification or detection systems. “We believe that AI integration with our customers will become more intimate, in that body language, gestures and the like will become control mechanisms – waving your hand at a door will open it, pinching in the direction of a window will close it, etc.”

He adds that AI will also become much more mobile, and our phones will become our AI platforms.

Over the next three years, Horrell expects to see the incorporation of historical data into CCTV analysis to become more widely deployable. “With the acceptance of cloud solutions, the processing of this data has become more cost-effective. CCTV trend analysis and the ability to alert on outliers or extraordinary scenarios will be more widely deployed. We also expect that cameras will seem to be able to ‘understand’ context and suspicious behaviour and alert on it. We do believe that it is moving in the direction of AGI where a single model can cater for multiple bespoke use cases, where this is not currently possible.”

Vermeulen says the next three years promise to be a defining era for AI. “We anticipate a future where AI completely revolutionises industries. Imagine fully automated security systems and retail experiences – that is the power of next-generation AI.”

Services on offer

Ending with the services the respondents offer customers today, Vermeulen says Secutel offers its SecuVue platform, packed with advanced AI functionalities (facial recognition, human pose detection, licence plate recognition, people counting, XFS events, etc.). “We are constantly innovating, with exciting new features planned for release in the coming year.”

“Iris AI is launching SnapGuard into the mainstream market in the next few months,” says Furter. Our biggest drive for the next year is to make the human posture analysis module, specifically the body language understanding feature, a standard feature in our technology. We will also finalise the research into our skeletal biometric recognition product.”

Horrell says that DeepAlert currently offers false alarm filtering, advanced camera health reports, and integrations with most video management systems. “We are working on integrating long language models into our architecture and enabling people to create bespoke alert configurations, i.e., people can create their own alerts for what interests them. We are also working on automated virtual guard tours and specific models for the detection of fire, PPE, loitering, and directional movement.”

For more information contact:

• DeepAlert, +27 21 201 7111, [email protected], www.deepalert.ai

• Iris AI, [email protected], www.irisai.co.za

Share this article:

Further reading: