Gun Detection Tech

tuckerj

Newbie
I would like to know what someone more seasoned in life than I am thinks of this development in technology and its application. Please help me out. Thank you. (cross-posting to facebook as well)
 

shoobe01

Established
Maybe they are just flagging it because he has his finger on the trigger? :)

I assumed this existed already. It's not "gun detection" it's just machine vision, and you can have it look for all sorts of stuff. You just have to decide you want to, or need to.

Hard for me to come up with broad based uses for this. If they expanded to stuff like gait analysis that could foolproofly detect concealed carry, that would be much more interesting. Maybe useful, but maybe worrying.
 
I'm not an expert in the field, but have some passing knowledge that might be interesting. This isn't meant to detract from what I think is actually a pretty cool project, but I wanted to put it into some context for those that might not have any background in computer vison or machine learning.

The framework for what hes doing here has been around for a while. It is likely that they are using a trained version of one of the existing open source image classification algorithms like AlexNet or YOLO to do the classification of gun vs no-gun. Like shoobe01 said, this could have been trained to detect any other arbitrary object just as easily. The detection performance is a function of input image quality, training data, and available computational power. So basically to see this concept come to fruition, you would be looking at relatively high resolution color security cameras to catch anything less obvious than a brandish at the camera. The video would then have to stream back to a box on site somewhere to do the processing. If we consider a school or some other relatively large building, that is a lot of different cameras that would all need to be processed simultaneously with as little latency as possible. Point is that there is a lot of scaling work that needs to be done to move from this proof of concept to anything remotely fieldable.

I did appreciate that they declined the request to detect concealed and holstered firearms. I'm not looking forward to a day that I am constantly irradiated while walking around town just to identify that I am excercising a right.

References if anyone wants to dive down the rabbit hole on image classification
AlexNet: https://towardsdatascience.com/alexnet-the-architecture-that-challenged-cnns-e406d5297951
YOLO: https://towardsdatascience.com/an-i...r-multi-object-detection-in-images-99cf240539
 

shoobe01

Established
Okay, your school point finally gave me a useful reason to detect a brandished firearm. There HAVE been a handful of videos we've all seen where the AS/AK gets all ready out of sight, but on camera. Having that go "maybe gun!" to an operator who then looks, agrees, then is able to sound the alert even a few seconds early might help a lot.

But... pretty rare cases. Would have to be somewhere it's super not good to have a gun, where no good guys do carry them, around so it doesn't do functionally false positives and: we incresingly suck at the whole human in the loop part, with untrained folks, who have no authority, or just deciding to save costs. If implemented I see this being abused sometimes and then totally misused so it doesn't help at all where needed. We'll be reviewing post-event video where the recording says "I think that's a gun."
 
Would have to be somewhere it's super not good to have a gun, where no good guys do carry them, around so it doesn't do functionally false positives and: we incresingly suck at the whole human in the loop part, with untrained folks, who have no authority, or just deciding to save costs. If implemented I see this being abused sometimes and then totally misused so it doesn't help at all where needed. We'll be reviewing post-event video where the recording says "I think that's a gun."

Agreed on all fronts. Theres also the opening for deliberate misuse. Imagine the potential chaos when some edgy but bright kid figures out what will happen if they pull up a gun video on their phone and put it infront of the camera. If there isn't a human operator reviewing the flagged images and the call is just routed to the local law enforcement, I can see things going very badly.
.
 
Top