printer friendly version

Smarter monitoring with AI

Over the past few years, Deep Data has brought artificial intelligence (AI) to the surveillance market, changing the way people monitor and manage their environments. Part of the service the company delivers alters the way remote monitoring is done, making better use of human resources by making better use of technology.

Hi-Tech Security Solutions asked DeepData CEO, Dr Jasper Horrell for more insight into the new world of smart monitoring.

Hi-Tech Security Solutions: We know that staring at a screen or screens is not the optimal way to monitor anything. The question is, what technologies are there that are reliable enough to run in the background and alert operators when something significant happens. Traditional video analytics partially fulfilled this function, but we still see many false alarms.

Horrell: In the past, manual video monitoring was possible, however, with an ever greater number of cameras deployed and streams demanding attention it has simply become an impossible task to monitor multiple streams accurately and consistently.

The first intelligent filter was line crossing and/or motion detection. This certainly reduced the need for consistent monitoring of video cameras. However, there are still many false positives when using only this system. DeepAlert makes use of both motion triggers and object detection using deep learning models.

Hi-Tech Security Solutions: What makes your solution different from others that are currently available? And why would your solution be more reliable than traditional video analytics or monitoring systems?

Horrell: The first filter of motion triggers is well established and easy to implement, however, it is of little use knowing that a line has been crossed or there is ‘something’ moving in a video stream. Often these alarms are triggered by foliage blowing in the wind, pets, or authorised people in the scene. Object identification is critical in reducing the number of false alarms.

This next, more challenging step, is to accurately identify the object that caused the initial trigger. Once the motion has triggered, our system sends a small portion of the video stream from the camera to the analyser, which resides in the cloud.

The system uses deep neural network technology to process inferences repeatedly and determine whether the video image matches known classes of objects. The DeepAlert system has been trained on a large number of images of classes of interest (30 different classes) that have been captured on the system for this purpose.

It is this deep learning process that ensures the accuracy of object detection, even under low light conditions and from practically any camera stream, including Infrared and thermal cameras.

Even though the DeepAlert system is extremely accurate, the system has an additional filter – an ‘alert confidence threshold’, which is user settable. Once the system has identified the object, it declares the certainty of that decision and alerts are sent out if the alert confidence threshold is exceeded. In sensitive scenes, a low alert confidence threshold may be set while in a very busy, less sensitive scene a higher alert confidence threshold may be set resulting in fewer alerts being received.

Hi-Tech Security Solutions: Is your system developed locally? How do you ensure it keeps up to date and takes ‘learnings’ from one system to update all the other systems using your solution?

Horrell: Yes, DeepAlert has been developed entirely locally. I [Dr Jasper Horrell] am the founder and CEO of Deep Data and have led the development team since inception in 2015. I was Head of Science, Computing and Innovation for South Africa’s participation (SKA SA, now called SARAO) in the international SKA project – the largest radio telescope in the world – and have applied my knowledge and expertise in deep learning models to video analytics.

DeepAlert is constantly evolving and the development team is very agile, listening to clients’ feedback and fine-tuning aspects of the system and adding functionality in response to client requests.

In general, the analytics takes place in the cloud so images from all sites are analysed by the same system which means that any changes and improvements apply across the whole system. The system also has a ‘train’ button that can be exercised by monitoring staff when an image appears incorrectly identified. This continuous training of the system improves detection even further over time.

Hi-Tech Security Solutions: Is your solution installed at a client’s location or in the cloud – or both? If cloud-based, how do you ensure you have enough bandwidth for video transmission etc.?

Horrell: Although the ‘heavy lifting’ (deep learning analytics) is done in the cloud, the system does not require a constant video stream to analyse the images and perform the object detection accurately, thus the bandwidth requirement is surprisingly low. We have flexible deployment options, but most current sites make use of the hybrid deployment with both an on-site component and the cloud component.

Hi-Tech Security Solutions: If someone wants to implement your solution, what do they need in terms of equipment? Do you have recommended edge hardware (cameras, for example) they need to use or can you integrate with most IP cameras out there?

Horrell: DeepAlert can integrate with practically any IP camera. There are three main deployment options – a hub on-premise which connects to the cloud for the analytics, a bespoke server (for enterprise-level installations and on-premise high-security requirements) and direct to the cloud for smaller or remote installations. The system only requires VGA-level resolution and can therefore work with many older surveillance systems.

The DeepAlert system also integrates with many video management systems so video monitoring staff can still operate off one unified user interface.

Hi-Tech Security Solutions: Do you sell your solution as a service, i.e. for a monthly fee, or does the client have to pay for licences (for a VMS etc.) as well?

Horrell: DeepAlert is sold using the Video Surveillance as a Service (VSaaS) model with a modest monthly fee per camera stream. There is no long-term lock-in and software updates are included in the monthly fee. An annual fee is also possible. (DeepAlert is sold through distributors in SA.)

Hi-Tech Security Solutions: Can you give examples of where your solution works, such as what environments and in which verticals?

Horrell: Perimeter surveillance and intrusion detection are the main use cases for DeepAlert. Our results show that DeepAlert reduces false alarms by over 95% when compared to motion-triggered alarm systems.

DeepAlert is also being used in the health and safety industry to monitor compliance, especially with respect to the wearing of PPE; recently, DeepAlert has developed a model for the detection of wearing face masks to assist with COVID-19 mitigation measures.

We are also moving into new AI analytics for the retail market with Unusual Behaviour Detection and People Counting.

Share this article:

Further reading: