As someone who has been involved in video surveillance for over 30 years, and specialising in external detection using video technology for the last 10, in both commercial and high security environments. I have been watching the development of artificial intelligence (deep learning) systems with more than a passing interest. What has led me to write this article is the huge spike in AI products over the last 12 months that are being promoted as the latest and best, with perfect detection and no false alarms. To good to be true right? well yes and no. I decided to carry out my own tests on a variety of popular AI only systems, from the free built in to the camera options, through to high end server systems including cloud and appliance based. Keeping an open mind I tested all the products in exactly the same way in a very controlled environment (the 4 cameras around my house) these cameras are all 1080p with a 4mm lens and all set to 25fps. One thing became very clear no matter what product was tested, I could defeat all of them in roughly the same way. If the AI system cannot see enough of the person then it simply does not alarm (think walking behind a wheelie bin or with an umbrella). I also found AI did not like people crawling or moving slowly in the shadows. The further away the object then the worse the systems performed. Anything beyond 45-50m was very hit and miss, typically I would expect decent detections with true video analytics down to 20 pixels but with AI you really need 50 pixels plus on target so your detection distance is greatly reduced.
If I walked across the detection area in a normal fashion in good light then it worked perfectly every time albeit slower than I would have liked. A typical detection time was >2 seconds which is much higher than leading video analytics solutions. This is why as far as I can see there are NO CPNI approved AI solutions. Are these issues a real risk? Well it depends what you are trying to achieve. If your just looking to be notified when the UPS man arrives or protect small well lit unobscured areas then its fine, but if you want to stop the professional intruder who is fully aware of how security systems work then probably not. AI systems are improving all the time but until they can detect accurately and reliably in poor lighting and with persons occluded by other objects, walking between cars for example and at longer distances then they are not suitable for medium to high security applications. The good things about AI is its very very good at false alarm rejection which makes it excellent at ignoring things like cobwebs, wildlife and reflections. Its also fantastic at identifying specific objects for example distinguishing a car from a bicycle or lorry which has traditionally been difficult. So in an ideal world you would have a combination of technologies, video analytics and AI working together utilising the best of both. There is such a product available 🙂 If you need more information about anything I have written then please feel free to drop me a line. Please don’t ask me to name the AI products tested as I won’t do that, its just not me, anything else is fine. Thanks for taking the time to read this and I hope you found it useful.
Stay Safe Ed Click to watch video.
Call on 01772 330 777 or Email firstname.lastname@example.org to speak to a specialist in video analytics and AI integration.