Perimeter Intrusion Detection Built for the Outdoors

Most perimeter analytics fail the same way. They alarm on rain, on headlights, on a plastic bag, on a fox, until the control room stops looking. The difference here is that an object is classified as a person or a vehicle before a rule is ever tested against it.

The Modules

Two Detectors for a Boundary

A perimeter is a rule about a place: this line is not to be crossed, this area is to stay empty, nobody should still be standing here after two minutes. IREX carries two modules that enforce rules of that shape, and the honest thing to say up front is which one belongs where. ObjectTrack Pro is the outdoor module. MotionTrack Pro is the quiet-indoor one. The product documentation recommends ObjectTrack Pro for a perimeter, and so do we.

ObjectTrack Pro runs a two-stage neural pipeline: a first network annotates objects of known classes and tracks them between sampled frames, a second decides what class each object is, and only then is the object tested against your line, your region or your dwell rule. Because the classifier runs before the rule, an event can be filtered to people, vehicles, or both, with a confidence threshold you set. That is what stops a night of weather from becoming a night of alarms, and it is why the module holds up in unfavorable outdoor conditions where a motion-led detector does not.

MotionTrack Pro starts from motion instead: it finds movement, segments the moving object against the object size you calibrated, then passes it to a network to be classified as a person or a vehicle and tracked against the rules. It reports motion in under a second and a classification in under three, and it carries the loitering rule. Its own documentation is blunt about the trade: it is insensitive to what kind of object moved, so it is recommended for quiet indoor environments and will generate an unacceptable number of false alarms outdoors. The same insensitivity makes it harder to spoof, which is why it still has a place inside a building.

The intrusion network was trained on hundreds of thousands of images and tested against public benchmarks including i-LIDS, the Imagery Library for Intelligent Detection Systems developed by a UK Home Office laboratory, and the detectors are in production on electrical power stations, oil terminals, pipelines and gas distribution facilities across several countries. IREX publishes no per-module accuracy figure: accuracy is measured on your own fence line during the pilot, against criteria agreed in writing, and the measured numbers go into the contract.

Read Why Our Intrusion Analytics Is Different

The Rules

What You Draw on the Frame

Every rule below is drawn or typed per camera and named by the operator, and the name you type becomes the event name the whole platform searches on. Nothing here is fixed in the model.

Line Crossing

A signal line drawn straight onto the frame, in either direction. Keep the tripwire off the edge of the picture: the minimum distance from its end points to the frame border is half an object width, or crossings at the edge are missed.

Region Entrance

A polygon rather than a line, for an area that should have nobody in it: a switchyard, a tank farm bund, a closed platform end, a right of way. Entry into the region is the event.

Motion in a Region

Movement where movement is itself the signal, on a plant room, a service corridor or a stretch of track. It is also the rule that carries safety-gear evaluation, which records helmet and vest compliance as an attribute on the event.

Loitering

A no-loitering zone plus a consider loitering time limit, in MotionTrack Pro. Somebody standing at a gate for four minutes is a different question from somebody walking past it, and this is the rule that tells them apart.

Off-Duty

The inverse alarm: a mandatory presence area with an allowed absence limit, so the event is the empty gatehouse rather than the intruder. It is how a guarded checkpoint is monitored without watching the guard.

Object Filters and Thresholds

People, vehicles or both, per rule, with a classifier threshold on the recognition probability. When vehicles are selected, type and color recognition can be switched on with them, so the event says what kind of vehicle crossed.

Saved Tracks

Stored motion trajectories for the people and vehicles in the scene, searchable in the media player afterwards. An intrusion becomes a route through the site rather than a single frame at the fence.

Image Stabilization

A digital image stabilizer keeps tracking usable while the camera is shaking, which on a mast in wind or on a fence-mounted bracket is the difference between a working channel and a source of noise.

Event and Camera Tags

Custom tags on events and on cameras, so a control room filters by north fence, inner perimeter or outer perimeter instead of by camera number. On an estate with hundreds of channels that is what makes the alarm list readable.

The Hard Parts

Range, Weather, and the Fence Itself

How Much Fence One Camera Actually Covers

There is no honest single answer in meters, because the constraint is not distance but pixels. The documented requirement is more than 50 pixels per meter inside the region of interest, a minimum object size of 32 × 32 pixels for a person and 45 × 45 for a vehicle, and a mount between 3 and 15 meters with tilt from +15° to +90°. Feed those into the lens and the mast height and you get your real coverage per channel, which is what a site survey is for. MotionTrack Pro will detect bare movement on as little as 8 × 8 pixels and more than 25 pixels per meter, but classification, the thing that suppresses false alarms, needs the larger figures.

See Camera Requirements

Night, Weather, and What Thermal Actually Does Here

Contrast is the variable that matters after dark: 20% object-to-background for classification, 10% for bare motion detection. Infrared illumination and thermal imagers are both used on perimeters for that reason, and the platform takes them, but be precise about how. A thermal imager registers as its own sensor type, with IP, port and camera number, and its own events, among them temperature, line crossing, intrusion, motion in region and object escape, arrive on the same screens and alarm monitors as camera events. The IREX analytics modules themselves are specified against visible-light video. Thermal widens what the estate sees; it is not a second input to the detector.

When the Fence Is the Thing That Changed

A cut mesh panel, a section pushed flat, a ladder left against the line, fresh spoil or excavation on the approach: these are conditions no trained intrusion module ships, and they are exactly what StreamVLM™ was built for. Damaged or broken fencing along the perimeter is one of the documented example prompts. A sentence becomes a detector on the cameras you choose, with its own confidence threshold and alert cooldown, and several prompt-defined detectors run on one channel alongside the trained module. Two limits to keep in view: StreamVLM scores sampled frames, so it reports a condition that is visible and that persists, not a subsurface one and not a fast-moving object; and it needs a GPU node. Shipping in beta on selected instances.

Explore StreamVLM™

Around the Detector

The Rest of What a Long Perimeter Needs

A fence line is rarely one camera near a control room. Five things decide whether perimeter analytics survives contact with a real estate.

  1. PTZ, under an Operator

    A PTZ camera registers with a port, optional ONVIF credentials and a speed multiplier, and an operator pans, tilts and zooms it from the video player to get a closer look after an alarm. Each use is logged as an event of its own. Be clear about the boundary: presets, tours, auto-tracking and slew-to-cue are not documented platform features, and the intrusion detectors themselves run on fixed cameras. The fixed camera raises the alarm; the operator drives the PTZ.

  2. Fence Sensors and Field Devices

    IREX unifies a camera estate with the field devices around it, and the platform architecture treats seismic and perimeter sensors as an input class alongside CCTV. What is documented as a named integration is narrower and worth stating plainly: Sigur for access control, Netping as an external device, ONVIF device and NVR registration, and a REST API. Beyond those, any alarm monitor can push its events with frames to any endpoint reachable over HTTP, and events arriving the other way drive the console. If your existing perimeter intrusion detection system speaks HTTP, it can be wired in both directions. A specific sensor integration is a scoping conversation, not an assumption.

  3. Monitors That Know the Hour

    An alarm monitor switches itself on and off on a schedule, with windows set by day of the week and by specific date. A yard that carries legitimate traffic at two in the afternoon is an alarm at two in the morning, on the same camera and the same rule. A manual override does not cancel the schedule.

  4. Edge Where the Link Is Thin

    Most perimeters are a long way from the control room and from good bandwidth. An edge server runs the analytics at the site and sends events rather than raw video, keeps recording when the link drops, and synchronizes automatically once it returns, which is what an unstaffed gatehouse or a pipeline right of way actually needs.

    Compare Deployment Models
  5. A Camera That Stopped Seeing

    On a perimeter measured in kilometers, a blind camera is an unmonitored kilometer. TamperTrack detects obstruction, defocus, repositioning and quality degradation per camera and alerts on it like any other event, which is how a long fence line stays honest between site visits.

    See Camera Tamper Detection

Requirements

What the Camera Has to Give It

Camera Mount
Fixed cameras only for the detectors. PTZ, mobile-phone uploads and drone footage are not supported inputs for ObjectTrack Pro or MotionTrack Pro.
Geometry
3 to 15 m height, tilt +15° to +90°, pan −75° to +75°, roll −5° to +5°, and more than 50 px/m in the region of interest for detection with classification.
Stream and Load
1280 × 720 to 2592 × 1944 at 10 to 30 fps. ObjectTrack Pro is rated high on computational complexity, MotionTrack Pro average, so the two size a deployment differently.

FAQ

Which module should we actually run on a fence line?

ObjectTrack Pro. The product documentation recommends it for outdoor perimeter work and states that it performs better than MotionTrack Pro because of its two-stage detection and classification design. MotionTrack Pro is documented as suitable for quiet indoor environments and is expected to produce an unacceptably high number of false alarms outdoors. Both are available and both are included; the recommendation is not a commercial one.

Will it detect someone crawling or rolling across the line?

That is what the Perimeter human detection mode is for. Since version 4.40 ObjectTrack Pro switches its tracked object types on separately, People, Vehicles and Abandoned items, each with its own settings, and the People type carries five tracking modes: People tracking, People tracking (top view), Safety gear detection, Pedestrian tracking, and Perimeter human detection. The last of those was added to capture people at a secured site including during unusual movement such as crawling or rolling, which is the posture an upright-person detector is most likely to miss on a fence line. The mode is chosen per camera in the module settings, so the channel watching the outer fence and the one watching the yard do not have to run the same one.

How do you keep false alarms down at night and in bad weather?

Structurally rather than by a filter bolted on afterwards. The object is classified as a person or a vehicle before any rule is tested, and each rule carries its own object-type filter and confidence threshold, so weather, foliage and light changes do not reach the alarm list as intrusions. A digital image stabilizer holds tracking together when the camera shakes. What remains is a placement problem: contrast has to reach 20% for classification and there has to be more than 50 pixels per meter where it matters.

What is the detection range?

There is no single number and we will not quote one. The documented constraints are pixels per meter in the region of interest, a minimum object size of 32 × 32 pixels for a person and 45 × 45 for a vehicle, and a mount height between 3 and 15 meters. Range therefore falls out of the lens, the sensor and the mast on each run of fence, which is what the site survey computes. Anyone quoting you a flat range figure is quoting a lens, not an algorithm.

Do you support thermal cameras?

Yes, as a registered sensor type. A thermal imager is registered with IP, port and camera number, and its own events, among them temperature, line crossing, intrusion, motion in region and object escape, arrive alongside camera events on the same screens and alarm monitors. The distinction that matters: IREX analytics modules are specified against visible-light video, so thermal extends what the estate detects rather than feeding the IREX detector. Infrared illumination on a visible-light camera is the other common answer on a dark perimeter.

Can an alarm point a PTZ camera at the target?

Not automatically, and we will not imply otherwise. What is documented is manual control: a PTZ camera is registered with a port, optional ONVIF credentials and a speed multiplier, an operator drives it from the video player, and each use is logged. Presets, tours, auto-tracking and slew-to-cue are not documented platform features. The intrusion detectors themselves run on fixed cameras.

Do you integrate with seismic sensors, buried cable, fence sensors or radar?

Not as named integrations, and the honest answer matters here. The platform architecture treats seismic and perimeter sensors as an input class alongside CCTV and access-control devices, and in the product what is documented is Sigur for access control, Netping as an external device, ONVIF device and NVR registration, a REST API, and an alarm monitor that can push events and frames to any endpoint reachable over HTTP. If your perimeter intrusion detection system accepts or emits an HTTP call, it can be wired to IREX in both directions, which is also how an intrusion event drives lights and sirens. A specific sensor integration is a scoping conversation, not something to assume from this page.

Can it detect tunneling or digging under the fence?

Only to the extent that something is visible on the surface. No IREX module detects subsurface activity, and a camera cannot. What a camera can see is spoil, excavation, disturbed ground, a vehicle stopped where none should be, equipment left against the line, or a damaged fence panel, and StreamVLM™ turns any of those into a detector from a plain-language prompt. Treat it as an early indicator for a patrol to check, not as a tunnel-detection system, and note that StreamVLM scores sampled frames and ships in beta on selected instances.

Does this run on its own or does someone have to watch it?

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.

Walk Your Fence Line with Us

Perimeter analytics is a placement problem before it is a software one. A site survey on one run of fence is worth more than a specification comparison.