We actually have a new server with a Tesla T4 which we recently purchased, though it definitely is not a 256GB RAM beast (yet...I think it has enough slots for that). We're currently testing a couple other things before we get back to turning it up but we will be using it for some deep learning tests very soon.
I agree completely - PPF/PPM results are highly dependent on other factors like environment, mounting height and light quality which are easy to overlook.
For our calculations, the majority of cameras we’ve worked with so far have higher FOVs ~90 degrees. Our current data shows we need roughly 17PPF/55PPM with a pixel area of ~500-1000 to make a detection, but we can detect at further distances with lower FOV cameras ~60 degrees.
The visible portion of the weapon needed to detect is approximate - reducing the visible portion of a smaller weapon at 35’(11m) has a larger impact than the same reduction on a larger weapon at 10’(3m)
Do you find the pixel area of the detection to be a more relatable metric than PPF/PPM?
We’re doing light quality testing this week - what other factors do you find valuable to test?
The problem of gun detection is not new and main issue on "regular" 1080p video security camera footage is lack of information to make good detection.
For example you are saying: "Our current data shows we need roughly 17PPF/55PPM with a pixel area of ~500-1000 to make a detection". From this follows that gun needs to be sqrt(500px)=22.4px or sqrt(1000px)=100px. This would be roughly between 1/2 and 2 metres in size for 55ppm. Unless i read this numbers wrongly. Handgun of 25cm length on 55ppm would be roughly 14pixels (55/4=13.75). And assuming you are using Yolo v3, image would be downsized and it will cause even more data loss.
So i am curious what size of the gun(handgun) in pixels custom Yolo version you have created, requires to make correct detection.
Interesting, is anyone aware of, or have any intel on Avigilon's gun detection analytics? I was just given a presentation from an Avigilon rep recently, showcasing essentially the same analytics (visual detection of a gun, triggering an event/alarm on a VMS), that they are developing, to be available sometime within the next year or 2.
From what I recall, their version took into account not only the shape/size of the potential weapon, but also relied on ques like, how a person was holding the weapon. Apparently their version runs on the camera side, thus requiring you to purchase their equipment and swap out any cameras you may already have.
It may have also had the ability to run on an appliance (assuming you had their VMS), but I can't recall, I kept nodding off at parts of the presentation where he would try to sell us on their VMS platform (it's amazing... we get it). Super nice guy though.
I've actually seen a demo of it, I have not installed it though. The demo, they held umbrellas and broomsticks like it was a gun, and the analytics didn't trip. Again, it was a demo environment, not real but it was impressive.
the real-world effectiveness of using this type of object-recognition analytic relies entirely on the premise/assumption made in their own video above at the :22 mark:
"That shooter is going to take out a weapon..."
Is this really commonplace in active shooter scenarios? What percentage conceal the firearm until they get inside?
i am not denying that the object recognition technology being used in this product works as described - nor that existing surveillance cameras can be used in conjunction with this analytic to recognize common firearms....
but how much of a difference can this approach actually provide in real world active shooter situations?
it seems like pretty much anyone with funding (and a decent back story) can use publicly available technology to make their own gun recognition analytic and bring their own active shooter solution to market... purportedly to 'save lives'.
and because we are all in the security industry and seek this very same goal, imo we are most susceptible to 'solutions' that are seemingly designed to solve exactly what we are all trying to do day in and day out... even with no empirical evidence that the 'solution' would make any actual difference in situations that the solution was purportedly designed to prevent.
"Is this really commonplace in active shooter scenarios? What percentage conceal the firearm until they get inside?"
Yes, it is commonplace for active shooters to take out a weapon before they shoot at people. In the case of Sandy Hook, Virginia Beach, Columbine, Virginia Tech, and Parkland as well as many others, the shooter exposed the weapon long before shots were fired. In many cases the shooter walked in with weapon exposed from parking lot.
"but how much of a difference can this approach actually provide in real world active shooter situations?"
The difference between first sight and first shot can be massive. There is an opportunity to prevent an entire event from occurring if an SRO/police is given intel in a timely manner. (picture of assailant, weapon, location, time) This information is critical when you are going into a chaotic & violent environment like an active shooter scenario. Turn cameras into force multipliers. Reducing time to get medical care to victims, stop bleeding, is critical.
"The difference between first sight and first shot can be massive. There is an opportunity to prevent an entire event from occurring if an SRO/police is given intel in a timely manner. "
The bolded sentence above contains the very weak qualifier 'can be' - in order to set up your statement that 'there is an opportunity to prevent' - which i have far less confidence in than you do.
the time span between first sight and first shot is far more likely to be measured in single digit seconds.... rendering the analytic useless because the SRO/police will not have the intel in a 'timely enough manner' to do anything to 'prevent'.
I think gun detection analytics are far more valuable in convenience store stick up prevention - as the weapon is displayed (and most often not even fired) in order to get the clerk to open the register.
But VC firms know that govt orgs will spend far more money on 'solutions' purporting to prevent active shooters than insurance companies will spend on solutions to prevent convenience store stick ups.
this is the first time that I have ever rebuked someone for voting one of my comments 'funny'.
the hawking of inferior solutions that have no actual value is not funny at all imo.
instead, I think this is just one of many examples of how humans are dumb and can easily be convinced to try out 'solutions' that appear to be tailored to emotional response vs solving any specific problem at all.
I don't care if you have an MBA or a military background, if you're a security vendor I expect you to have the sense to not use "Let's Encrypt" certificates on your website. That's like buying used East German body armor on E-Bay. I'm sorry but if your cloud stuff is gonna call 911 I get to expect it to have a super pristine internet presence.
It shows very poor information management hygiene. It demonstrates an inappropriately lack of respect for the concept of identity management. It suggests the dev team behind their product has low crypto clue (yeah I just said low crypto clue about a Navy Seal, I'm not happy with that either.) It suggests they don't look at their whole internet presence as a cyber target. It suggests they do things the old way where you barked at a summer intern in the marketing department to create a web site in five minutes on zero dollars budget and no you can't have 10 dollars to buy a real cert. "That's not us that's the marketing department" does not cut it any more.
And I just posted this minutes ago so we both should stand down and give the vendor a chance to get a word in edgewise ;-)
We should be very critical of these types of autonomous active shooter detection systems and I believe they deserve the same oversight as fire alarm systems. Without that oversight, they are just another tool that may help us during the worst. Because this system relies on visibility of a firearm, I doubt it will ever be 100% accurate and will just remain yet another tool in our layers of security, at best.
Many K-12 buyers may look at this as one of those "better than nothing" solutions, but I don't see how this really mitigates the issue of someone shooting at you, but I don't have the unique perspective of a combat SEAL.
I wouldn't spend public funds on this tech. We do use cameras with built-in gunshot detection analytics, because it is included with the cameras. Just another tool that we should never rely on.
What can you rely on during an active shooter event? Humans are good at detecting gunshots, especially multiple gunshots with an erratic cadence. People will call 911 in a hurry and the call takers will get very good and actionable information for their responders. Train and equip your people to quickly snap into action when hearing gunfire.
Well now that this is capability is public knowledge, future shooters will just know to conceal their firearm until it's time for the bodies to hit the floor. Shouldn't need an analytic to let you know what's going on at that point.
I agree that this should be treated as an accessory rather than as a first-line defensive measure.
These mass killers have adapted their tactics over time to increase the victim count. It's predictable that they will better conceal their firearms until time for the gunfire, with or without ZeroEyes.
Assuming these mass killers would behave in the way "normal" people behave when statistics from past shootings show a majority of mass killers actually take their weapon out, sometimes in the parking lots of schools well before they get in the door. While I agree eventually most cameras will have gunshot detection, for now adding a few exterior cameras onto ZeroEyes provides a rare chance to get ahead of a shooting instead of minimizing the damage during one.