Skip to main content

BEHAVIOR ANALYTICS

Crowd, Loitering & Behavior Analytics

A person box and a car box are just pixels. Sentinel turns them into behavioral signal — how many people are in a zone right now, who has lingered too long, which way the gate count is moving, and when a known subject breaks their own established routine. Every rule is deterministic and explainable. No black-box LLM deciding what matters.

Command Center heatmap showing geographic crowd density across the camera estateClick to enlarge

Geographic crowd density across the whole estate, live.

120s
Default dwell threshold
per-zone loitering trigger, fully configurable
Per-zone
Live people count
warning + critical capacity bands
1.2M
Events scale-tested
zone aggregation on covering indexes
0
Black-box decisions
every alert is a rule you can read

ZONE INTELLIGENCE

Six deterministic behaviors, drawn on your own footprint

You draw the zones — a polygon over a lobby, a polyline across a gate, a perimeter around a dock. Sentinel counts COCO person and vehicle detections against those shapes on the agent, in real time, and emits an event only when a rule you configured is crossed. No model is guessing intent; the math is auditable line by line.

CROWD DENSITY

Per-zone people counting

The agent runs COCO person detection and counts everyone inside each zone polygon live. Set warning and critical capacity thresholds per zone; an event fires the moment a band is crossed, and utilisation renders in the Command Center crowd-density widget.

LOITERING

Dwell-time detection

Each zone carries its own dwell threshold (default 120 seconds, configurable). When a tracked person stays past it, Sentinel emits one loitering event per episode — with start time, end time, and max dwell — and bumps the threat dial.

LINE CROSSING

Directional entry / exit counts

Configure a zone as a polyline with a direction — entry, exit, or both. A person centroid crossing the line emits a directional event, giving you accurate gate and corridor counts without a turnstile.

MOTION

Motion detection + AI gating

Frame-diff motion detection runs on the agent with per-camera sensitivity. It doubles as a compute gate: when a scene is stationary, expensive AI inference is skipped entirely — so the CPU spends cycles only where something is actually happening.

SMART FILTERING

What to detect

A surveillance default allow-set drops junk classes — chairs, plants, TVs — at the source, before they ever become an event. Per camera you can narrow further to people-only or vehicles-and-plates, keeping the operator feed clean and the database lean.

PATTERN-OF-LIFE

Deviation watches

A subject who breaks their own established routine triggers a watch. Sentinel scores each person’s regularity and fires a deviation alert when their observed gap runs well past expected — the single biggest "told me something I did not already know" moment.

CROWD & DWELL

Density and loitering, counted against the zones you draw

Counting happens on the agent, against polygons you place over the scene — so the same camera can watch a lobby for crowding and a back corridor for loiterers at once, each with its own thresholds.

  • Per-zone live people count from COCO person detection, with configurable warning and critical capacity bands rendered as utilisation in the crowd-density widget.
  • Loitering tracks person detections inside a zone over time and fires one event per person per episode — start, end, and max dwell — at a default 120s threshold you can tune per zone.
  • Line-crossing zones count directional entry and exit at gates and corridors from a simple centroid-crosses-the-line rule.
  • Threshold crossings drive the alert sound and the threat dial, and scale-tested aggregation across 1.2M events keeps the dashboards responsive.
See the AI vision pipeline
Live

Every scene counted live — people and vehicles, against the zones you draw.

PATTERN-OF-LIFE · DEVIATION

When a routine breaks, Sentinel raises its hand — honestly

Deviation alerts build on a Shannon-entropy regularity score over each subject's sightings. The gates are deliberately conservative: Sentinel only claims someone is off-pattern when it has enough history to make that claim credible.

  • Alerts fire only for subjects with a regularity score of 60 or higher and at least 30 sightings in 90 days — below either floor the person is silently skipped rather than over-claimed.
  • A yellow alert at 2.5× the expected gap, escalating to red at 5×; alerts auto-resolve the moment the subject is sighted again.
  • Operators can subscribe to a VIP, POI, or witness to watch them under relaxed gates — every subscribe, pause, resume, and removal is audit-logged with the operator and reason.
  • No autonomous action: deviation alerts surface the signal and the operator decides. No auto-enrolment, no auto-pan, no notification cascade.
Explore pattern-of-life
Pattern of Life
last 90 days · 339 sightings
MonTueWedThuFriSatSun
0006121823
Peak: Mon 09:00 · 11 sightingsRegularity 78 / 100 · Highly regular

A 7×24 regularity baseline per subject — deviation alerts fire only when the pattern is strong enough to trust.

WHY IT HOLDS UP

Explainable by design, accountable by default

Behavioral analytics is only useful if you can defend it. Sentinel applies the same honesty contract everywhere: confidence-aware filtering, gate-aware empty states, and a full audit trail on every operator decision.

No black-box intent guessing

Every behavior here is a deterministic rule over counts, dwell timers, line crossings, and a published regularity score — not an LLM inferring what a person "meant" to do. You can read exactly why any event fired.

Confidence-aware math

A single MATCH_CONFIDENCE_THRESHOLD gates the behavioral math: low-confidence matches are excluded and plate-only cameras are kept out of face-pattern analytics, so deviation signals are built on data you can stand behind.

GOVERNANCE

Audited & lawful by default

Deviation subscriptions capture the operator, an optional policy justification, and an auto-expiry; every state change is audit-logged. Behavioral monitoring runs against the platform’s lawful-basis and audit rules.

Learn more

See behavior analytics on your cameras

Draw a zone, set a dwell threshold, and watch raw detections become signal. Request a vetted live demo and we will respond within one business day.