The big jump of performance of Generative AI (artificial intelligence) demonstrated by chatbots such as ChatGPT4 is certainly a top technology fanfare in 2023. These chatbots are large language machine learning models. The writer was previously involved in image (or convolutional) machine learning to improve the performance of video surveillance cameras. Has image AI advanced as much as language AI? The short answer is nowhere close. ChatGPT is like an avalanche. AI security camera is like a peaceful river. This article aims to reveal the gains of a peaceful river over the last 5 years.
Everything has a sweet spot which is up for each person to define. For video surveillance cameras, this article defined a camera with 4MP (megapixels) as the sweet spot 5 years ago (2019) and one with 6MP as the current sweet spot (2024). This definition implies that the retail pricing of a typical 6MP camera at present is about the same as 4MP five years ago. Now let us change the context to AI. AI may mean machine training with large data sets or fixed rule-based computing without training. Is the current sweet spot security camera more AI-capable than the last generation?
The current model can do human and vehicle detection crossing a virtue line or entering a virtue area to trigger recording. This means that the system engineer can preset a virtual line or a virtual area on the camera image, and footage recording will take place when human or vehicle is detected by the camera to have entered the line or area. Is this a big deal? Yes, it is for security because this recording has a context of intrusion. The last generation could trigger recording when an object entered the line or area. The object can be a cat, a fallen leave, or the shade of a moving cloud. We can easily infer that the last generation camera would record a lot of footage and only a small percentage pertains to human or vehicle. Not only storage utilization has improved, the effort of retrieving useful footage for police evidence has sped up.
How applicable is the above statement? This article refers to professional grade video surveillance cameras commercially in mass production. Some related technical background information can be found on www.cnz.co.nz. Go to the Resources Tab and read Technology Snippets.
The image processor handling AI is a piece of hardware inside the camera. A software application in the PC which receives camera streams can do AI as well. Compucon spent 5 years to develop AI capabilities in 2014 to 2019 timeframe using post-graduate students from Auckland based universities. The Compucon AI computer model was able to interpret camera footage such as “a small girl wearing a backpack is walking down the pathway”. We can easily tell that the context of this AI output is more useful than human entering an area. Our team reached 65% accuracy in our tests on a fixed project site. We may think 65% accuracy is not bad as an academic effort. In reality, the Police for example will not accept anything that has a failure rate as high as 35% except that it can be used as an excuse for not being able to solve a criminal case. We abandoned the project in 2020 when our R&D budget dried up for AI development.
AI does have a place in security cameras. Do we need to wait another 5 years to see more significant AI empowerment?
Download this document:
Please note: To download this document, you require to be logged in as a SUBSCRIBER.