Not motion detection.
Intent detection.
Every security camera system detects motion. Verkada detects motion. Avigilon detects motion. RMADOR goes further — building a continuous picture of what each person or vehicle is actually doing, assigning a 0–100 threat score, and acting before a threat reaches your door.
Six layers of intelligence
Each layer builds on the last. By the time a threat score is generated, the AI has assessed every camera, tracked every subject, and evaluated dozens of behavioural signals.
Layer 1 — Perception (RF-DETR, 50–150ms)
Every camera feed is analysed continuously by RF-DETR, our fine-tuned detection model. People, vehicles, drones, bikes, animals, and packages are identified the moment they appear — across every connected camera simultaneously. 7 object classes, 60+ fps on the server GPU.
Layer 2 — Tracking (ByteTrack + DeepSORT, 10–25ms)
Each detected object is assigned a persistent track ID across frames. The AI builds a continuous record of where each person or vehicle came from, how fast they're moving, velocity vectors, and where they're heading — even across multiple camera handoffs.
Layer 3 — Trajectory Analysis (10–20ms)
Speed, acceleration, approach angle, dwell time, and zone entry/exit events are computed for every tracked subject. A vehicle accelerating toward a gate, or a person repeatedly pausing at a fence line, is flagged the moment the pattern is established.
Layer 4 — Temporal Behaviour (LSTM + Transformer, 20–50ms)
LSTM and Temporal Transformer models classify what a subject is doing over time: NORMAL_TRANSIT, LOITERING, APPROACH, CIRCLING, STATIONARY, RETREAT, or AGGRESSIVE_APPROACH. Intent is inferred from the sequence of movements, not a single frame.
Layer 5 — Behaviour Graph (GNN + CEP, 10–30ms)
A Graph Neural Network analyses patterns across multiple subjects simultaneously — convoy approach, tailgating within 1.5m of an access event, coordinated probing, or object abandonment. Single-subject analysis cannot catch coordinated threats; this layer does.
Layer 6 — Bayesian Threat Scoring (5–10ms)
Every signal from every layer is weighted and combined into a single 0–100 threat score, adjusted for object type (drone weighted highest at 0.30), zone sensitivity (restricted zone multiplier ×1.0), and time of day (00:00–05:00 ×1.4). This score drives every alert and automated response.
One number. Every signal.
Every behavioural signal feeds a single 0–100 threat score. Your alert threshold is configurable per zone — a restricted area triggers at a lower score than a public entrance. Night-time automatically increases sensitivity.
What the AI watches for
These are the behavioural patterns that real security incidents share. Each one contributes to the threat score independently — and several together produce an immediate alert.
Loitering
Stationary in a zone for more than 2 minutes
Often precedes an opportunistic intrusion or scouting activity.
Repeated approach
Returns to the same zone 3 or more times in an hour
Indicates deliberate surveillance or probing of your perimeter.
Unattended object
Item left behind after a person moves away
Package abandonment is a primary indicator in threat assessment protocols.
Coordinated entry
Multiple vehicles arriving in close formation
Associated with organised forced-entry attempts.
Tailgating
Person follows immediately behind an authorised entry
The most common way to defeat access control without triggering alarms.
Aggressive approach
Fast movement aimed directly at a barrier or entrance
High-speed vehicle threats require a sub-3-second response.
Perimeter probing
Slow movement along a fence line with repeated pauses
Classic reconnaissance behaviour ahead of a planned breach.
Camera evasion
Subject consistently keeps their back toward cameras
Deliberate avoidance of identification — significant threat indicator.
See it detect a real threat
Our live demo runs a full incident scenario on your dashboard — from first detection through bollard raise — in real time.
Request a demo