MarshallAI x Avarn: Security Operators' view on deep learning enabled security monitoring
AVARN Security has been our partner in the security industry since 2017
In collaboration with Avarn and their clients we have built a complete solution suite for modern security monitoring that leverages recent advances of AI and deep learning. In this post Timo Räihä, Head of AI Solutions at Avarn Security shares his experiences with MarshallAI and reveals his thoughts about the future of security monitoring in the era of AI.
I’ve been doing security and video based supervision 15 years now. During this journey the technology has advanced a lot. Coverage of video monitoring has increased due to lowered price tags per unit. On the other hand the quality of video streams has improved from 240x240 analogue to 4K digital 360° videos. Digital VMS has given the ability to remotely adjust configurations of cameras and speed up scanning of footage afterwards. Some surveillance camera providers have also started to provide some aggregated analytics functionalities as well as other smart functionalities. However, my experience is still that they are under development and thus require a fair bit of configuration to deliver moderate results.
All this development has made my and my colleagues’ everyday work easier, no doubt.
I’m interested in sharing my experience with the next generation surveillance and mass monitoring system Marshall AI. MarshallAI runs on existing camera infrastructure (you don’t have to buy new cameras nor new licenses), it integrates into your workflows and most importantly it works off-the-self in many use cases such as:
People analytics: Reporting statistics to real estate owners: People count, Gender, Age, Areas, Direction, Time
Face recognition: Finding the wanted person for police
Tracking: Waiting for signal if a person of interest continues his/her journey after sitting for an hour in a coffee shop
Abandoned luggage: Especially in airports it’s great to be alerted immediately if someone leaves a luggage behind.
Area monitoring: At certain entrances there shouldn’t be any movement between 10pm and 5am
PPE Monitoring: Related to the area monitoring, especially in the transportation hubs where there is reconstruction work ongoing, with MarshallAI I can easily monitor whether people entering the construction area are wearing personal protective equipment, and to easily identify and restrict passengers access to these areas.
Beyond the ones mentioned above I’ve used the MarshallAI platform for vehicle counting and tracking, traffic (indoor & outdoor) detection, detecting abandoned luggage and visible hand guns or weapons in areas where they are heavily prohibited, such as shopping centers, transportation hubs, sports stadiums and schools.
One more thing worth mentioning is that MarshallAI solutions are designed to meet the highest privacy standards. Faces of individuals, except persons of interest, are automatically anonymised and the data won’t under any circumstances get uploaded to a public cloud, because it’s running within our and our customers’ on-premises.
Future
These features have really disrupted my everyday work. But the thing I want to discuss is the future role of security supervisors and how it will change.
I’d summarize this transformation as “from supervisor to metavisor”. The main task is shifting from monitoring people and their behaviour to fine-tuning and supervising an automated solution, MarshallAI, to perform ever better…
- Get filtered information, solve difficult edge/ambivalent cases -> focus on what matters
- Train MarshallAI to do more work for you -> focus on what matters
- Use MarshallAI to monitor thousands of surveillance streams simultaneously
My typical shift starts with going through MarshallAI’s pickups and recommendations of hot shots / relevant events/alerts during last 24 hours. MarshallAI also reminds me if something special is is going to happen today or if I have forgotten to do something. Then it’s usually time for the magic. In the past I would have started with the first video circuit. - Now I focus on other more productive tasks until MarshallAI flags something!
Instead I’ll start training my data-driven co-pilot. Typically, our daily training is fine-tuning the model performance by going through edge cases (where the model wasn’t very sure about a detection) which MarshallAI has identified itself, or some of the operators has flagged that the filtered alert provided by MarshallAI wasn’t helpful for some reason. I’ll check these edge case samples and provide the model with the correct information, what it should have detected from the sample or what its reaction for the detection should have been.
