AI is the buzzword of the moment, a silver bullet that will resolve all challenges – but will it? What are the limits of its accuracy? Where is it best applied?
AI video analytics has been around for some time. Most models are primarily based on object detection (usually people or vehicles); when these objects are detected in the field of view, an alert is generated. The main benefit of this application of AI is the ability to filter the many nuisance alerts being triggered on cameras and NVRs based on line crossing, motion detection or intrusion detection.
Even at a relatively modest scale, such as residential estates, monitoring requires some form of filtering as the basic triggers from cameras still generate too many alerts for operators to address effectively.
The challenges for an AI model to accurately detect people or vehicles in the field of view are similar to those of the naked eye. The smaller the object (further in the distance), the more difficult it is to detect. If the object is partially obscured (a person hiding behind a wall), this too presents a significant challenge to AI models. A lack of contrast (usually dark clothing worn in a poorly illuminated area) also makes detection difficult. Bright lights inadvertently directed at the camera, usually intended to illuminate the field of view of another camera, can 'washout' any image and make detection almost impossible.
Certain weather conditions, such as rain, can sometimes confuse AI models, causing nuisance alerts, much like spiderwebs on the camera lens.
The cornerstone of any surveillance system is the physical installation. A well-designed and installed camera system ensures that the field of view of cameras are well illuminated, overlap to some degree, and that the cameras are operating well within the manufacturer’s specifications. This will give your AI video analytics model the best chance of detection and multiple chances of detection of a particular incident.
As with any surveillance installation, a stress test, which includes the AI video analytics model, should be conducted to verify the efficacy of the AI model in conjunction with the physical installation. Some configuration may be required to ensure optimal accuracy.
Most cloud-based AI video analytics models can easily be plugged into existing surveillance systems. Most service providers allow a testing period before estates must commit to purchase. Additionally, most cloud-based services are sold using the ‘software as a service model’, which usually means licences per camera, per month.
Ideally, it is best to ‘try before buy’! You are invited to submit footage or images (MP4 or jpeg format) of a stress test, or even an incident that you have experienced where there was an intrusion of people or vehicles at your estate or an estate where you provide security services. Simply upload your details and your footage/images at https://www.deepalert.ai/smart-estates-conference/ and stand to win a prize in the lucky draw!
| Tel: | +27 21 201 7111 |
| Email: | mark@deepalert.ai |
| www: | www.deepalert.ai |
| Articles: | More information and articles about DeepAlert |
© Technews Publishing (Pty) Ltd. | All Rights Reserved.
printer friendly version