Interest in the subject of using a CCTV camera as an alarm device is rising.
At Gensix Technology we feel quite qualified to discuss this topic, as we implemented these solutions as far back as 2020, and have plenty of feedback on the efficacy of such systems.
Before reading through the whole article to draw a conclusion – let’s give you our view right off the bat.
Currently we don’t think this is a viable solution.
Way back in 2007 at an IEEE Conference on Advanced Video and Signal Based Surveillance, one paper concluded the following:
“While the political/commercial drive for promoting and the willingness of society at large to accept such technologies generally exist, the actual gap between users’ expectation for intelligent actionable and/or searchable video data and the reality that the state-of-the-art technology solutions can offer is yet to be bridged.”
In 2022, we feel that this conclusion remains relevant.
While we have deployed Dahua CCTV Cameras and Recorders – there is very little difference between the top brands and the solutions as we describe them.
MOTION DETECTION
STANDARD MOTION DETECTION
For a start this discussion is only increasing because of the use of Artificial Intelligence (AI) in CCTV video analytics.
Standard motion detection systems simply monitor the difference of pixels between consecutive video frames. If there is enough change between frames, the system determines a motion event.
These systems have no reliable means of determining what actually caused the change in pixels. They cannot identify the nature of the specific motion event.
Such systems produce an insane number of alerts for all sorts of insignificant reasons – leaves blowing in the wind, cloud shadows moving across the lawn, changes in light levels, people, dogs, cats, and monkeys will all activate a motion detection event.
Obviously such a system would be ineffective as a critical alarm device.
Not only would it produce hundreds of false alarms, but it really annoys the occupants and neighbours. They will switch it off or just ignore it – which is hard to do when it starts beeping in the middle of the night.
SMART MOTION DETECTION
Manufactures have claimed – over the past few years, to have employed artificial intelligence algorithms for video analytics that improve the accuracy of motion detection to 98/99%
Let’s just comment on this 98% claim. Indeed 98% of the time the system will detect a human, or a vehicle -whichever is set in the software. But 2% of the time, it will false alarm. (I am using the manufacturers numbers – but really, our experience has been a much higher rate of false alarm than what they claim)
Let’s assume that their claims are accurate. How appealing do you find being woken up twice in the middle of the night by blaring sirens – to find that a bug flew in front of your camera?
Bugs are pretty common at night, in the presence of a light or Infra-red source.
FALSE ALARMS AND MISSED DETECTION
Generally we have found these systems false alarm when bugs are in front of the camera, rain drops fall directly in front of the camera, or there is a sudden light change in the scene that involves the entire image.
Often we have been able to fine tune the settings in these cases.
It can take a lot of time and site visits over many months to get it right though. Getting the system to ignore these issues, also often results in reduced capability to actually detect what it should.
With one particular manufacturer we found problems result when there is a parked vehicle in the scene. If there is any form of motion in that scene – the system detects motion and a vehicle, and then alarms as a moving vehicle – (Even while the CCTV system was not set to detect vehicles). The entire detection platform is rendered useless when there is a vehicle parked in the scene.
Now, what about the systems failure to detect a human?
We had a case where the system did not detect intruders who stood in front of the camera, and stole laptops from an office – through the window.
Now this is not particularly a system failure. This failure identifies a different problem.
The target must be in full view of the camera – the whole body works best. The camera should not be at an angle with the target – it works best when the camera observes the target in a straight line, and the movement of the target must at least exceed some set threshold. If the target moves slightly, pauses, and moves in this way – as thieves usually do – there is a reduced probability of detection.
Customers expect that the camera will detect a human in it’s field of view.
Many more cameras are required for detection though when you are limited to about a 12m detection range – base this on the DORI distance required for recognition level surveillance. Also trying to accommodate straight in-line scene observation. There is a significant increase in the number of cameras required to properly detect a human. Yet – you still have the false alarm problem.
Other problems that we experience are related to the cameras.
The CCTV manufacturers are using low technology solutions to provide video and power to their cameras. Both video and power often share a common ground. Ground loops are prevalent, and these voltage irregularities tend to false activate the event detection of the system.
THE FUTURE OF AI
Until now, the CCTV camera has proven ineffective as an alarm device.
The only potential benefit at this stage, is when a CCTV camera is linked to another motion detection device. Once the device is activated, the camera can be accessed to determine the source and validity of the alarm event. This trend in visual alarm systems is known as video alarm verification.
Outdoor motion detectors, often experience high false alarm rates – but we have found that the better quality devices, are better at motion detection than the cctv cameras.
Many manufacturers are abandoning the common HD Analogue CCTV camera system and recorder – because AI requires heavy bandwidth and compute capacity.
These HD systems have reached their sell by date, when it comes to their capacity to meet the demands of modern smart video systems.
The new trend is IP (Network) CCTV cameras.
One driving factor is what is called Intelligence at the edge, or distributed video surveillance.
In the edge model, all the video processing and AI application is done at a single camera, for a single video stream. This eliminates the high powered compute capacity that is required by a single recorder, to process multiple video streams via it’s cpu.
A single camera is equipped with a cpu that is sufficient for processing a single video stream. The pre-processed data is then sent to the storage device, reducing the compute load required from a central recorder or attached server.
This model also provides for more advanced and demanding AI algorithms to be applied to the video stream.
CAN IP CAMERAS BE USED AS ALARM DEVICES
Currently, we have to say no.
These systems are also around 60% more expensive than a comparable HD analogue system – but in our opinion, still cannot be deployed as a stand alone alarm device.
Although these cameras have upgraded and updated AI algorithmns, they still operate on the same principles, and reflect similar levels of accuracy.
While I cannot confirm this, I believe that Dahua is using a customised form of open source software known as open-cv. Actual tests show a 91.07% degree of precision, and a 68.77% recall measure for dahua’s object detection algorithm at 0.6 Threshold.
An open source system like oiyolo shows a 93.23% degree of precision, and 71.32% recall. Studies have shown YOLO outperforms all other platforms – but still isn’t perfect.
Another problem with these neural networks is that they are not most effective in a real-time system , or where any degree of speed is involved in the scene.
Our final conclusion is that it will be some time yet, before a CCTV camera with AI – can reliably be deployed as a stand-alone alarm system.
These systems have vastly improved the standard motion detection feature. Many of our customers use these systems in combination with push notifications on their phones with much success, having a number of successful preventions under the belt.
But – they are not ready to be fitted to a loud alarm device, or an armed response monitored alarm system – yet!
If you are interested in AI enabled CCTV recorders , for the home – you can find our available range at : Shop – AI CCTV Recorders
