Detection, notification, response. In an active-shooter incident those three factors decide the outcome: how fast it is detected, how fast the right people are notified, and how fast they respond. A camera nobody is watching helps with none of them. Weapon detection video analytics on the cameras you already own takes care of the first two and gives the third a head start.
The Problem
Most schools, campuses, and public buildings already have security cameras, and they are already meant to provide situational awareness. The gap is that no administrator or staff member can watch them continuously. Detection therefore usually begins with a person realizing what is happening, and notification begins after that.
IREX closes the detection and notification gap. GunTrack handles visual weapon detection, and AudioTrack listens on the same estate for a gunshot, breaking glass, a scream, or a loud sound it cannot place. Audio needs no new hardware: the module runs on any camera that already has a microphone. Detections route in seconds to the staff, law enforcement, or dispatch you nominate, through real-time crime centers, 911 workflows, and the Sover secure messenger.
A firearm is usually the smallest and least legible object in the frame, so GunTrack reads each candidate twice: a detection pass across every frame of the stream, then a verification pass on the same region at the camera’s maximum resolution. The two passes are set out below, and they are what stands an umbrella down and what still reports a weapon that has been repainted, cut down, or carried muzzle-down against dark clothing.
Weapon detection is combined with watchlist face recognition and person re-identification, so once a subject is identified the platform can follow them through corridors and across cameras even if clothing changes, giving responders real-time location rather than a last-known sighting.
The Chain
Visual weapon detection runs continuously on every connected camera, so a weapon that is drawn or carried into view is reported before the first shot, which is the moment an acoustic sensor first hears anything. Gunshot, breaking-glass, scream, and loud-sound detection run alongside on every camera with a microphone. None of it depends on anyone watching a monitor.
The alert reaches nominated staff, security, and law enforcement at the same moment, on the alarm monitor and in a Sover channel on the phones they already carry, with the frame, the camera, and the map location attached, rather than traveling through a call chain.
Responders get live tracking across cameras, indoors on floor plans, so they arrive knowing where the subject is now.
Response Time
In a typical mass shooting incident, police are notified after the shots have been fired. At this point, there has already been a delay. Once victims realize what’s going on and process the situation sometimes the first call is to family members. Calls to the police still have to go through a routing system for information gathering and then are finally dispatched. According to the FBI the average response time for law enforcement is 3 min. So where can we save time and essentially save lives?
Every step in that chain is human, and each one spends time the incident does not give back. The FBI’s figure describes the last step only: three minutes is measured from the moment law enforcement is notified, not from the moment the weapon appeared. Camera-based detection moves the start of the clock. GunTrack reports the weapon when it is drawn or carried into view, before a shot, where an acoustic sensor can only hear the first one, and the alert leaves for the people who will respond in the same second, with the frame attached.
That is where the time is saved: the gap between a weapon appearing and a qualified person knowing about it. Detection and notification collapse into one event, and the response clock starts earlier, with the location already known. Nothing responds autonomously. The alert is a signal for a person to verify, and every verification, decision, and assignment is written to the audit trail against a Case ID.
“As a law enforcement officer with over 16 years of training and experience as a police detective, major incident commander, police technology instructor, and supervisor, I understand that with IREX early detection reduces response time and will save lives. If the shooter is in the school. IREX will track the suspect throughout the hallways regardless of clothing changes. Police will also have real-time information about the location so they can quickly respond to the correct door. In a chaotic situation, accurate real-time information helps law enforcement stop the killing.”
“Unfortunately tragedies on campuses have increased tremendously. Mental health is on the rise due to the current world climate. School districts really need to take a look at themselves, and I ask if they are in business to change children’s lives. To do so, they need to be alive to be given that chance. IREX is that chance, technology is a must and if you agree then use the best. IREX Real-Time proactive approach allows law-enforcement and school officials to act immediately. IREX clearly saves lives, you cannot put a price on a child’s life, or anyone for that matter. IREX is the leader and should be used by every school district in the United States. Expect the unexpected, don’t let the next tragedy hit home.”
“Are you tired of seeing young kids murdered at schools by active shooters? No matter what your political affiliation is, we can all agree that people with intentions to kill innocent children should be stopped. IREX AI uses the most advanced, state-of-the-art technology to detect crimes in real-time to stop predators in their tracks. IREX AI connects cameras and sensors to a secure, private cloud, analyzing data to assess threats in real-time. This provides vital, pro-active opportunities to stop active shooters while ethically protecting the privacy and American liberties of citizens.”
IREX Clearly Saves Lives
Notification
A detection that lands on a screen nobody is watching is not a safety system. The event that opens on the alarm monitor is pushed into a Sover channel at the same moment, so the officer on foot, the duty officer at home, and the dispatcher are looking at the same frame in the same second.
Switch on messenger notifications for an alarm monitor and the monitor gets a chat of its own in Sover, IREX’s end-to-end encrypted messenger, with the authorized users already in it. Every Gun detected event that lands on the monitor arrives there as the snapshot, the event name, the camera, the location, and a link that opens the playback in the IREX player, on Windows, macOS, Linux, Android, iOS, or a browser. Members comment, forward the alert, and upload photos and video from the scene, so the channel becomes the record of what the team did about the alarm and not only the record that it fired. Alerts are organized by channel rather than by inbox: a user joins, mutes, and leaves the channels that match their duties, which is what keeps a busy estate from turning into an unread pile of email and SMS. The same events still reach the real-time crime center and the 911 and dispatch workflows; the messenger adds the responder who is not at a desk. Sover runs on the same private cloud as the video platform, so nothing about a weapon alert leaves the infrastructure you control.
See What IREX and Sover Do Together →
Two Passes
One look at a gun-shaped object is not enough to page a police officer. GunTrack reads every candidate twice, at two different resolutions, before an event exists at all.
A convolutional network trained on firearms analyzes every frame of the incoming video at the standard resolution the camera streams for analytics. Where it finds a weapon it builds an object box around it and calculates a similarity score. The network is trained on the two size classes that matter operationally, handguns and long guns, from datasets built from mass-shooter simulations, publicly available police and military footage, and human-reviewed synthetic composites of pistols, revolvers, rifles, and shotguns placed onto real CCTV frames. This pass has to be cheap enough to run without pause on every connected camera, because detection that depends on someone watching is the gap this module exists to close.
The candidate is then re-read at the highest resolution the camera can provide, on the small region the first pass isolated rather than on the whole scene. That is where a firearm is separated from an umbrella, a phone, or a printed image of a gun, because the answer is in the detail of the object and not in its outline. The pass also reads the dynamics of the weapon and the person carrying it, so a stationary object, a poster or a screen showing a firearm, does not become an event. Version 4.38 replaced both halves of this pipeline, the detection and the classification models, and the gain landed exactly where the second look matters most: weapons further away from the camera. That is also why IREX added GPU nodes to its Texas instance in 2024: a larger network at the second pass reads a smaller object.
An event fires only when three conditions hold together: the object box sits inside the region of interest you drew, the similarity score clears the threshold you set, and the weapon persists across a set number of consecutive frames. All three are per-camera settings, tuned on your own cameras during the pilot. The resulting Gun detected event carries Critical priority by default and lands on the alarm monitor of the roles that own that camera.
Hard Cases
Most of the published discussion runs one way: not raising an alarm on an umbrella, a phone in a raised hand, a poster or a screen showing a firearm, or a toy carried openly. IREX does claim that direction, and analyzing the context rather than the shape alone is how the module separates benign objects from genuine threats. The other direction matters just as much and gets skipped in datasheets. A real weapon repainted in bright colors, wrapped, cut down, slung muzzle-down against dark clothing, or lying half inside an open bag is still a weapon, and its silhouette is exactly the one a shape-only detector waves through. The training pipeline randomizes the color and finish of every composited weapon for precisely this reason, so an unusual frame color or coating does not exempt an object from the first pass. The verification pass exists for both directions, because in both of them the answer is in the detail of the object rather than in its outline.
At a large public research university in the United States, the platform alerted campus security to a concealed firearm within the first ninety days of operation. Security reached the subject in forty-five seconds, de-escalated, and established that the item was a realistic replica. The university has since extended the deployment across every building and parking structure. Read the sequence carefully, because it is the design working rather than failing: the module reported a gun-shaped object in a place where guns are forbidden, and a person established what it actually was. A platform that only ever raised an alarm on confirmed live firearms would have stayed silent on a realistic replica being carried through a campus, which is not the outcome any safety office wants.
A weapon detector that cries wolf gets switched off within a month. The 2023 rework that introduced the second pass was reported at the time as cutting false alerts by more than tenfold, and every detector since carries its own confidence threshold and alert cooldown. Alarm monitors scope events by camera and type to the roles that own them, and thresholds are tuned on your own cameras during the pilot. Where real incident footage does not exist and could not decently be collected, the model is strengthened with human-reviewed synthetic data: roughly three in four generated samples are rejected before a reviewer passes about three hundred verified examples per weapon class into training. Placement carries as much of the result as tuning does: GunTrack performance depends on how large a weapon appears in frame, so a camera position is chosen for the balance between the area it surveys and the pixel size of the object it has to read.
AudioTrack
A weapon is only visible if it is in frame. Sound is not, which is why the gunshot class stays on this page as one leg of the response chain, while the acoustic module as a whole has its own page. AudioTrack is a detector, so it layers onto a camera that is already running GunTrack: the same channel, the same event list, the same alarm monitor.
Around the Weapon
The firearm is often the hardest thing in the frame to see, and the people around it are not. StreamVLM™ prompt-defined detectors add those conditions as a sentence an operator types, rather than a module somebody has to train, so an event still fires when the weapon itself is behind a counter, in a pocket, or out of view. StreamVLM™ is shipping in beta on selected instances.
“Detect people with hands up or lying on the floor.” Two tellers with both hands raised, two customers face-down on the tile, one sentence covering both. The platform reports the posture, not the crime: an event carries the frame, the camera and the time to whoever the branch nominates, and a person opens the clip and decides what to call it.
See How Prompt Detectors Are Written →
The same sentence, the other posture. From this angle the handgun is a dark object held low against dark clothing at the back of a bright hall, the hardest read in the frame. Six people flat on the floor are not. Run alongside GunTrack and AudioTrack, that gives the same incident three independent ways to raise an alarm.
From the Record
The module has a public paper trail. The detector launched in a city pilot in 2021, was rebuilt around the second pass in 2023, moved onto GPU nodes for distance in 2024, and in 2026 gained a hand-reviewed synthetic training pipeline and a place in a Texas school district’s safety program. The releases are here in their own words.
Optimizing Weapon Detection: The Role of Human-Reviewed Synthetic Datasets AI-powered weapon detection relies on massive training datasets. Real-life footage, however, is limited, sensitive, and often problematic. Building an AI that detects weapons… Read More → 
HorizonIQ Expands IREX Ethical AI Capabilities for Enhanced Weapon Detection in Texas and California Dallas, TX, USA: June 28th, 2024 HorizonIQ has bolstered the IREX Ethical AI bare metal instance in Texas by integrating new GPU nodes. These nodes now form part of a private cloud, delivering real-time weapon detection video analytics services across campuses, malls, and airport… Read More → 
Enhancing Public Safety: IREX Revamps Weapon Detector to Minimize False Alerts by Over 90% In our continuous effort to bolster public safety, IREX proudly unveils an upgraded version of its weapon detection video analytics module. This enhancement remarkably diminishes false alerts by more than tenfold, significantly elevating the efficacy of monitoring systems. Conseq… Read More → 
IREX Launches its New AI-Powered Weapon Detector in a Iconic City Alexandria, VA - July 26, 2021 – IREX announced today that it has started a pilot project with its new AI-powered weapon detector in one of the world's largest cities. With the new weapon detector, IREX allows cities to combat 21st-century problems such as mass shootings in… Read More → 
How Do IREX Cobots Reduce Response Time and Prevent Catastrophic Mistakes in Public Safety? The uncertainty in situational awareness causes the most dramatic mistakes during a public safety crisis such as mass shootings, terror attacks, or hostile protests. Fast and efficient communication is the key to overcome the "fog of war" and support high-stakes decisions. IREX S… Read More → 
The Three Pillars of Trust in Gun & Weapon Detection AI Modern cities rely on layers of technology to protect public spaces. Among them, gun detection video analytics stands out as one of the most critical advancements for security infrastructure. The… Read More → None, and deliberately. IREX writing policy forbids "guaranteed detection" or "100%" wording and requires any accuracy claim to cite its basis. The wiki holds no published per-module figure, so the commitment we make is procedural: accuracy is benchmarked on your own cameras during the pilot, against criteria agreed in writing beforehand, and the measured numbers go into the contract.
Yes, if the weapon is in view of a camera. That is the difference between visual detection and an acoustic gunshot sensor: the sensor starts the clock at the first shot, while GunTrack reports a weapon that is drawn or carried into view before one. The two are complementary, which is why AudioTrack runs alongside it for the weapon that is fired out of frame.
Because the two jobs need different things. Finding a candidate has to happen on every frame of every camera continuously, which is only affordable at the resolution the camera streams for analytics. Deciding whether that candidate is a firearm needs detail, not coverage, so the second pass re-reads only the region the first one isolated, at the highest resolution the camera can give. Version 4.38 updated both the detection and the classification models, and weapons further from the camera were where the improvement showed.
It analyzes the context and the detail of the object rather than the outline alone, which is how benign objects such as umbrellas or toys are separated from genuine threats, and that separation is what the verification pass is for. It is not a screening machine and it does not adjudicate legality: it reports a weapon-shaped object where weapons are forbidden, and a person decides what it is. At a large public research university in the United States that is exactly what happened, and the item turned out to be a realistic replica. A detector that stayed silent on realistic replicas would be the worse product.
No. This is camera-based visual and acoustic detection, so it sees what a camera can see. Concealed-weapon screening is a different technology class and IREX does not claim it. What the platform can do where a weapon is out of view is detect the conditions around it: StreamVLM™ prompt detectors for people with hands raised or lying on the floor, and AudioTrack for a gunshot, a scream, or breaking glass.
No. AudioTrack runs on any camera that already has a microphone, so there is no acoustic array to procure, mount, power, or maintain, and no separate sensor network to run alongside the cameras. The audio arrives on the camera stream as AAC, G.711 mu-law, or PCM, and it is optional per channel, so you turn it on where sound is worth listening to and leave the rest of the estate as it is.
IREX publishes no range, accuracy, or latency figure for acoustic detection, and we are not going to invent one for a website. The commitment is the same one the visual detector carries: we measure it on your own cameras during the pilot, against criteria agreed in writing beforehand, and the measured numbers go into the contract. Two things we can say now. Detection works from the microphone in the camera, so coverage follows your camera plan rather than a sensor grid. And it is not recommended in noisy outdoor locations such as underground stations and railway terminals, where constant ambient noise works against the threshold.
Yes, and that is the normal configuration. Several video analytics modules and detectors run on the same camera at the same time, and AudioTrack is a detector, so GunTrack and AudioTrack on one channel is supported. The two then cover each other: a weapon fired out of frame is still a sound that was heard, and a weapon carried in silence is still an object that was seen.
Whoever you nominate. Alerts route to real-time crime centers, 911 and dispatch workflows, and the Sover secure messenger on the phones staff and officers already carry, all with the triggering frame, the camera, the location, and a playback link attached. Each alarm monitor can have a Sover channel of its own with the authorized users in it, so an officer follows the channels that match their duties and mutes the rest.
In the reference deployments IREX publishes: Southern California law enforcement in Oceanside and El Cajon, where weapons in public spaces sit alongside traffic and watchlist analytics; Dublin Airport, at access points; the Peru National Police and the Lima and Miraflores command centers; and the Kazakhstan national smart-city program, which runs the full module catalog. In 2024 the Texas instance gained GPU nodes serving weapon detection across campuses, malls, and airports in Texas and California.
Often. Any camera that supports ONVIF or streams RTSP with H.264 or H.265 can be connected, but weapon detection needs a clear, higher-resolution view of the scene, so the camera requirements and placement both matter, which is what the site survey checks camera by camera.
GunTrack runs on CPU, but with limited performance, so IREX strongly recommends a GPU node for weapon detection. The GPU nodes added to the Texas instance in 2024 are what allowed larger networks and better reads at a distance, while CPU inference stays supported for customers whose data centers cannot take GPUs. Most other modules run at full performance on standard CPUs through the proprietary Synet inference framework, and IREX engineering sizes the CPU and GPU mix per deployment against camera count, module mix, and the servers you already own.
It runs on its own and alerts a person. Detections are signals for a human to verify: no response runs autonomously, and the verification and decision are logged against the same Case ID as the detection.
Keep Reading
Weapon detection is mostly a placement problem. An hour on site tells you more than a datasheet.