Gun detection analytic startups are increasing as the promise of AI and the threats of active shooters grow.
One company, Actuate (formerly Aegis AI), is being led by a Marine Corps Veteran Sonny Tai, who aims to use his experience in the military to help solve this problem.
We spoke to Aegis, to better understand their gun detection first approach, go-to-market, and how they position their video analytics in campus environments.
... that emergency calls to police take 5 minutes from the first shots fired
I find it very hard to believe that it takes 5 minutes for the first 911 call to reach police. Police are usually on-scene within 5 minutes of the first shots, but that is a lifetime when being shot at.
I wish I had a good answer to this question. I first heard of the "5 minute" statistic when I attended a "Run, Hide, Fight" training taught by a career LEO. I've then seen it used by multiple secondary sources, without being able to track down a primary source.
I've also used this statistic when speaking with LEOs ranging from patrol officer to Chief (I met a bunch of Chiefs at Axon Accelerate last week), and nobody has refuted it.
If anybody has a primary source that analyzes the time between 911 calls, that would be great. I know at Parkland, it was 45 seconds after the first round was discharged, which may suggest that 5 minutes is off, but I just don't have enough data.
The (5) minute number is referenced in a number of the civilian response training. The cause for the delay is referred to as “normalcy bias”. Amanda Ripley describes it in her book Unthinkable.
Many active killer trainers, including myself, do what we do to combat normalcy bias. I believe and hope the influx of related training has changed the delay and has saved lives, but I have no data.
I cannot find any contemporary study that lists the average timeframe of the first 911 call from the scene. Individual news articles and some case studies have calls coming in within a minute of the first shots fired.
I'm not defending or denying what has been said, just looking for data.
Brandon, you make a great point about combating normalcy bias. I think you can summarize your point with this statement... "you will not rise to the occasion you will fall to your level of training". Stress inoculation is critical if you want to react effectively during a critical incident. If you only train until you get it right you will fail under pressure you must train until you can't get it wrong!
why use videos of convenience store/bank robbers as examples of how their gun detection analytics work - then claim that this isn't what they intend them to be used for?
I find this interesting - because in strong-arm robbery situations (like the ones they used in their own demos), I think their solution would actually provide some value - because authorities can actually be alerted before anyone gets shot.
the display of the gun gets those being robbed to acquiesce and hand over the loot. and they aint calling the police while under that gun, are they? gun detection analytics could certainly help in these scenarios.
Plus convenience store/bank robberies are quite common.
My guess is that they figured out the 2nd paragraph above early on - and made videos to show what they could do... but then when VC became involved these investors wanted to focus on active shooter solutions instead - cuz that's what the buzz is all about now - and they shifted focus for the potential of a higher return.
except active shooters generally (I am no expert, but this is my perception) enter and begin blazing away -- rendering gun detector analytics somewhat useless - at least in comparison to the obvious value rendered in strong-arm takeover scenarios.
I'm one of the founders. Happy to answer any questions to the best of my ability.
why use videos of convenience store/bank robbers as examples of how their gun detection analytics work...
We used convenience store robberies for two reasons:
- As you noted, it is much more abundant, which means that with some very dedicated and creative scraping, it can serve as a valuable repository of initial training data.
- We also didn't want to traumatize viewers with video footage of people being shot in a demo video, so using store robberies was a good way to demonstrate our capabilities using real life security camera crime footage instead of stock footage.
My guess is that they figured out the 2nd paragraph above early on - and made videos to show what they could do... but then when VC became involved these investors wanted to focus on active shooter solutions instead...
Violent gun threats was our focus from day 1. It stems from my personal upbringing and passion about the issue, but you're right, market demand also trends that way. Whereas institutions would be willing to invest resources in security solutions to save lives during active shooter situations, stores are usually insured anyway and are unlikely to see a solution like this as worth the cost of investment, that is, until GPU compute becomes so cheap that you can essentially offer this for a fraction of the cost.
except active shooters generally (I am no expert, but this is my perception) enter and begin blazing away...
You are absolutely right. The value of computer vision gun detection is much less about prevention, but it's about the ability to mitigate issues with information asymmetry in real time. For example - in Parkland, law enforcement had no idea what the shooter even looked like until 7m30s after the first round was fired (https://www.sun-sentinel.com/news/103472606-132.html). Computer vision threat detection can provide real time, visual intelligence on:
- Location/Direction of Movement
- Armament/Equipment
- Appearance
...which are all critical pieces of information that enables law enforcement to neutralize the threat as soon as possible and for building occupants to make decisions that maximizes their chances for survival.
I know there have been a lot of companies who have come and gone promising computer vision gun recognition capabilities. I'm not saying we have it all figured out - the last thing we want to do is overpromise and underdeliver. However, we have a hell of a data science team, customers who love what we're doing for them, and we have some pretty distinguished investors who believe in us. We're striving to be different.
Thanks for the feedback! I've been an IPVM reader for over a year now, big fan of the community here!
Thank you! I think that if we put together a passionate and talented team and conduct ourselves with integrity, we can ultimately make an impact in a meaningful way that saves lives.
Your post was not provocative at all :-) It's natural to be skeptical, as you should be. There are too many vendors in this industry looking to make a quick buck but massively underdeliver.
I think there are ultimately two quick ways to tell if a company is serious in this space:
1. View their LinkedIn page and see if they have anyone with data science expertise (more specifically - experience with neural networks and computer vision). Anybody can train an off-the-shelf model on an open source data set and get it to work semi-well. It takes a lot more to make it reliable.
2. Have they raised significant venture funding? We have raised around $2M (will announce soon), most of which is from VCs who were early investors in companies such as Lyft, Palantir, and ClassPass - and they do not mess around with their technical due diligence.
Does that mean you don't currently notify police directly? (ie, the system sends SMS/Email to a list of users, and those people are expected to notify police)
At this point, we don't, as for now there are too many bureaucratic layers to work through to build out integrations with every law enforcement department.
However, we have plans for incorporating this in the future, either through:
(a. Integrating into a computer aided dispatch system after the alert has been confirmed by a human
(b Building our own monitoring center and calling 911.
We haven't prioritized this yet, because customers have actually asked to get the alerts themselves. Currently, the alerts go to 3-5 members of a school leadership team in education deployments, and the Director of Security and the SOC (if applicable) in corporate deployments.
We build integrations with VMSs as we onboard customers who use a specific VMS. We would love to work with Ocularis/Qognify, but we don't have a Qognify customer yet. Ironically, both my co-founder and I are University of Chicago alums and they use Qognify, but we haven't been able to land them as a customer yet :-P
AI for gun detection is something that is of interest to me. We do a lot of work in the K-12 space and are very focused on integration of systems to detect, lockdown and providing notifications and communications. This is about the 4th company that I have come across now but the pricing model is very reasonable.
The one thing I have noticed is that these companies are trying to find thier way and how do you go to market. If the vertical is going to be small business, retail this is a different sale then dealing with a public school that has bidding requirements or they need a state contract or a legal procurement path to sell to these clients.
Hi there, thanks for your questions! These are issues we often discuss with our customers:
1. We pull images at moderate resolutions and low frames per second. As most of our customers' cameras are limited by lens quality, this doesn't substantially decrease image quality, but substantially reduced bandwidth requirements. An initial deployment on 20-30 cameras is about equivalent to one person streaming Netflix. Also, on most VMSes we can leverage the basic analytics they're already doing, and only pull images when motion is detected, further reducing bandwidth by 80-90%. This means that bandwidth isn't a huge concern, and is definitely far cheaper than installation costly compute hardware on site.
2. There is actually very little delay across the internet. In all of our deployments we can receive and analyze an image within a second of it being stored by the VMS. The lag we see at customers is between their cameras and their VMS, which is a local configuration problem that we recommend customers work with their Security Integrator to address.
3. Most of our customers don't have bandwidth concerns, but for the ones that do we can use on-site motion detection (as in #1 above), or limited-scale deployments to address their bandwidth concerns. This is nicely counter-cyclical: When there is a lot of activity in hallways, usually people aren't using the internet as much, as they're leaving work or going to lunch, so our system fits nicely into existing bandwidth demand patterns.
4. Our system was build with privacy in mind from the ground up. Firstly, we don't analyze faces or any other personally identifiable information: We're just storing images without any relevant metadata. Second, we wipe all customer data every day, so the data retention time for even those un-annotated images is very short. Third, we use the data-transmission protocols specified by the customer's VMS, which their IT team has already approved as part of the VMS purchasing process. None of our code touches the customer's network except for an alert being sent back in the case of a detection. While we understand that some customers may still have concerns about any data being transmitted externally at all, we believe our system is the most privacy-compliant real-time Video Analytics solution currently available, and other solutions will require additional PII analysis, longer data retention periods, or more 3rd-party software or hardware touching a network.
1. In the article it was mentioned that you need 10-20 FPS , this isn’t low bandwidth, and I am sure you can’t detect weapon at 1-2 FPS. second I am not sure you can in the majority of the cases register to an RTSP stream and determine the amount of frames that you unless you configure camera by camera or at the VMS level.
2. I don’t understand how you can relay on the motion detection analytics. First, it is not reliable and produce massive amount of false alerts. Second, it is irrelevant in an environment which has constant motion which is the majority of your use-case as far as I understood it.
3. I would argue that there is latency issue within the internet especially for a critical service as you provide.
Consider the above I am not convinced what you presenting can scale.
Btw, what is the minimum object size that is required in order to determine that this is a weapon?
After careful analysis of the company’s market positioning, Actuate leadership decided to adopt the new brand name in alignment with its new features, which expand the firm’s offerings beyond gun detection.
The new features include intruder- and threat-detection AI solutions. The Actuate system can now alert customers to unauthorized entry to customer facilities and catch individuals acting in a threatening manner even before weapons are fully visible.
By broadening their offering Actuate should be able to address additional users' needs if the new offerings work. However, the move could also indicate that opportunities for a pure gun detection video analytics service are not enough to build a company around.
Rebranding, while not common, is not unheard of in the analytics market (e.g. Ever AI to Paravision, ImageIntelligence to Nirovision).
can Sonny tell us more about how these new features work in conjunction with the gun detection analytics?
I like the move to identify potential risk before a gun is ever displayed but I'd like to understand just how the new stuff and the old stuff - together - adds value.
ftr: hat tip to Sonny for snagging the .ai domain extension.