Artificial Intelligence (AI) is transforming every industry it touches, from healthcare, to transportation to agriculture. So, it should come as no surprise that the security industry is starting to be transformed through AI as well. In fact, security, automotive and healthcare are in the top tier of industry verticals to benefit from this amazing technology. Security has some major influences driving the adoption of AI, including the increasing threat of terrorism and more sophisticated security perpetrators.
If this technology scares you a bit, then read on – this article will help you understand AI in a bit more detail, why the impact for security will be so profound and why it is your friend rather than something to be concerned about.
AI is being incorporated into a wide range of different technologies. Unfortunately, while many software and hardware companies advise that their product utilises AI, they have simply added a few more ‘smarts’ to their software and claim it is AI ready to go. In light of this, the following will also help you identify true AI capabilities.
A standout capability of AI is its ability to eliminate mundane and boring tasks. Imagine you had the job of sifting through a photo taken every 10 metres of every road in Australia and your job was to recognise where street signs were placed (or those that had gone missing), as well as the type of road barriers and any pot holes. You would soon become rather bored. AI-enabled computers do not get bored and do not need to be supplied with food and water (other than power), so they will happily process every photo, and at a much faster rate than humans. The person previously assigned to this task can then perform more stimulating activities.
AI will take jobs, but only certain types of jobs, so it is more about an overall change of the job mix. Gartner, a leading analyst company, are predicting that in the US alone, AI will create 2.3 million jobs in 2020 while eliminating 1.8 million and that by 2022, one in five workers engaged in mostly non-routine tasks will rely on AI to do a job.
The quality of life for people around the world will increase as AI can support doctors to recognise illness and disease which may have been previously missed. X-Rays or MRI scans, for example, may be double-checked by an AI engine which understands what a good knee should look like, thereby letting doctors know if a patient’s is less than optimal.
If you need another example to demonstrate why AI is transformational and beneficial, think of the farmer who may have previously sprayed his entire crop with weed killer/insecticide to eliminate weeds and bugs, which is costly and adds unwanted chemicals to food. Using AI, this same farmer can now find just the weeds and bugs by using cameras monitored by AI to deliver a dose of weed/bug killer only to the target area, without impacting the entire crop.
To better understand AI and its possible applications in security, it is necessary to break it down into its various components. Tractica, a specialist research company in the AI area, have put together a list of the main AI functions in order of anticipated worldwide revenue size. I have added a column to indicate if this component will impact security – most are a ‘yes’.
# | Area | Security Relevant |
1 | Static image recognition, classification and tagging | Yes |
2 | Algorithmic trading strategy performance improvement | No |
3 | Efficient, scalable processing of patient data | No |
4 | Predictive maintenance | Yes |
5 | Object identification, detection, classification, tracking | Yes |
6 | Text query of images | Maybe |
7 | Automated geophysical feature detection | Yes |
8 | Content detection and classification – avoidance, navigation | Yes |
9 | Prevention against cybersecurity threats | Yes |
How does AI Work?
This article will focus on the areas from the above table offering the biggest level of value to the security industry, which are areas 1 and 5. A subcomponent of AI, which deals directly with image recognition, is a technology called Deep Learning. From this point onwards, this term will be used rather than AI as it is the most relevant of all AI capabilities for the security industry.
Why is Deep Learning so important? Think of facial, behaviour, gesture, licence plate and weapon recognition as an example, which all use a form of image recognition from still images or video.
Deep Learning solutions typically outperform traditional recognition technologies, which is one of the main reasons why the level of interest is so high. In fact, when trained properly, Deep Learning has now reached the point where its capability to recognise an object or person is equivalent to or better than that of a human.
The AI Knowledge Journey
In the following example (image 1), there is obviously an orange and an apple. If I was a software developer and I had to develop some software code to tell the difference between these two kinds of fruit, I could simply ask the computer to work out the colours of each fruit and say the job is done.
Image 1
However, in image 2, it is exactly the same fruit but this time the colour on my camera was not working and so the two different types of fruit are now black and white. The logic I so proudly used to distinguish between the two types of fruit in image 1 is no longer relevant as I have no colours other than white, black and shades of grey. To address this problem, I will need to add more logic to describe the fruit without the use of colour. Where do I start? Is the orange more round than the apple? Perhaps that can be my logic now, albeit not the best logic, particularly as some oranges are not perfectly round.
Image 2
Image 3 brings additional challenges as I have now added a new piece of fruit, which has both green and orange in it. My existing fruit differentiation logic is now almost worthless, and the recognition process is becoming more and more complex, which often leads to accuracy issues.
Image 3
This time, I want to write some logic to tell the difference between a dog and a mop, as per image 4. You would think this would be easy.
Image 4
But what happens if my dog looks like this (image 5)?
Image 5
All of these issues, plus many others, are addressed by Deep Learning. Rather than writing computer code to tell the difference between the orange and apple, this time I am going to provide a Deep Learning training computer with 50–100 pictures of oranges in different orientations and different lighting conditions. The Deep Learning engine utilises something called a neural network, with typically five to six layers of intelligence, with each layer responsible for recognising something about the image. The first layer might be responsible for recognising squares, rectangles and triangles before passing onto the next layer which performs another task and so on.
Rather than comparing against a reference image, Deep Learning builds a mathematical model of the item it has been trained on. As a result, it is essentially recognising something by utilising the mathematical representation of that item. The fact that it is a mathematical model is very important for the following reasons:
Image 6
- Mathematical models can be utilised to calculate with a high degree of precision how well an item appearing in an image or video matches the stored representation of that image. In the case of the orange example in image 6, the Deep Learning engine obviously matched the orange on the left with a high level of accuracy. Because the orange on the right has a lot of mould, the Deep Learning model still believes it is an orange, but this time the confidence level has dropped. In this case, additional reference images of oranges with mould could be added and the Deep Learning system would then improve on its capability to recognise oranges even with mould. This confidence-based approach enables organisations to only trigger alerts when a high level of confidence exists, thereby reducing false positive event alerts.
- Rather than utilising the Central Processing Unit (CPU) of a computer to calculate if something is an orange with 25 frames of video being sent to it, researchers have found that a Graphical Processor Unit (GPU) with many hundreds of cores can process Deep Learning mathematics at a much higher rate than a CPU. The rise of powerful GPUs has brought Deep Learning to life, as computers have not had the processing power in the past to process this type of workload. GPUs are enhancing other areas of the security industry as they also enable Visual Management System (VMS) platforms to process a higher number of video streams than a CPU only system.
Lastly, the final concept to understand: Deep Learning actually emulates the way that the human brain recognises something. It does this through the use of neural networks which are a series of layers in the mathematical model, with each layer being responsible for recognising a specific aspect of the image. There are various neural network models, with some simply passing the image mathematical from one layer to another and others involving a looping process between some layers to enhance accuracy.
To ensure your software provider is utilising Deep Learning, it is best to ask them if their software requires at least one GPU to run and if they utilise neural networks. There are some solutions which utilise neural networks without GPU, which are either not actually utilising Deep Learning or will most likely be slow.
In a previous article published in this magazine, I wrote about the arrival of Security 4.0, which is the next evolution of the security industry. GPUs and Deep Learning have enabled the industry to evolve into this next phase of security, particularly with the rise of off-the-shelf Deep Learning solutions like Briefcam. Deep Learning enables security to be event driven rather than observational, where the Deep Learning engine trigger is informing a security operator of an event he or she needs to be aware of rather than relying on the security officer’s ability to make the observation for himself.
Further to this, because of the large number of recognition technologies available, security can choose multiple ways to recognise a person or a vehicle and the like, correlated together. As an example, with the rise of people printing their own vehicle licence plates as a copy from another vehicle with a 3D printer, organisations will need to utilise more than one technology to determine the correct vehicle.
As a real-life example of the use of Deep Learning, albeit this time for marketing, a Melbourne-based company called Deep Recognition utilised Deep Learning to recognise the make and model of cars being driven on the road. Rather than trying to describe the vehicle to the computer system, they used a database of many different vehicles which are added into the Deep Learning training engine, typically being 60–100 images of each vehicle, so that it can learn what a Nissan Ultima and many other vehicles look like.
I hope this article has provided you with an easy way to understand AI and Deep Learning and why this technology is set to be so transformational to the security industry.