LOOKOUT
NVR
Sign In Download Free
AI Detection & Smart Alerts

AI that runs on
your hardware, not theirs.

Motion is recorded first. Detection names the person, vehicle, or animal, and it never decides whether the recording exists. It runs on the computer you own. A graphics card helps and is not required.

Try AI Detection Free How Alerts Work
What detection does
It names the event. It does not gate the recording.
Detection runs while something in the picture is moving. A quiet hallway costs less CPU than a windy tree line. Size that on the sizing page. Detection speed on a graphics card has not been measured.
Motion first
Every motion session becomes an event. Detection refines the name. A missed name does not delete the footage.
Person, vehicle, animal
Those are the classes the product is built around. You can give one a name so it sticks the next time.
CPU or a card
Windows uses DirectML when a card is present, and the CPU when it is not. Linux with no NVIDIA card uses the smaller image. A card nobody has probed counts as no card.
Detection Classes
What LookoutNVR can see.
Person, vehicle, and animal are the categories the recorder is built around. Other classes the model can see are there for filters. A name you give one object sticks.
Person
Detects individuals in frame. Separate from motion detection - triggers only when a person is actually present, not shadows or swaying trees.
Vehicle
Cars, trucks, motorcycles, bicycles, buses, and trains. Know when your driveway or parking area has a new arrival without checking footage.
Animal
Cats, dogs, birds, horses, cows, sheep, bears, deer, and more. Filter them out of security alerts or create dedicated wildlife monitoring rules.
Full Object Catalog
All 80 detectable object classes.
Every class can be used in per-camera filters and alert rules. Whitelist only what you care about, or blacklist what you don't.
People
person
Vehicles
car truck motorcycle bicycle bus train airplane boat
Animals
bird cat dog horse sheep cow elephant bear zebra giraffe
Outdoor & Street
traffic light fire hydrant stop sign parking meter bench
Accessories & Bags
backpack umbrella handbag suitcase tie
Sports & Recreation
frisbee skis snowboard sports ball kite baseball bat baseball glove skateboard surfboard tennis racket
Kitchen & Food
bottle wine glass cup fork knife spoon bowl banana apple sandwich orange broccoli carrot hot dog pizza donut cake
Indoor & Furniture
chair couch potted plant bed dining table toilet sink refrigerator microwave oven toaster
Electronics & Personal
tv laptop mouse remote keyboard cell phone book clock vase scissors teddy bear hair drier toothbrush

Detection runs on the computer you own. A graphics card is optional. The sizing page is where a camera count comes from.

Per-Camera Filters
Every camera sees different things.
Filter accordingly.
Each camera has independent class filters. Whitelist only the classes you care about, blacklist the ones that cause noise, or use a preset to get started in one click.
Security Preset
People & vehicles only
Whitelist: person, car, truck, motorcycle, bicycle, bus. Everything else is ignored. Best for driveways, entrances, and parking areas.
person car truck motorcycle bicycle bus
Farm Preset
Ignore common animals
Blacklist: bird, cat, dog, cow, horse, sheep, bear, zebra, giraffe, elephant. Alert on people and vehicles. Best for barn and field cameras.
bird cat dog cow horse sheep
Wildlife Preset
Animals only
Blacklist: person, car, truck, motorcycle, bicycle, bus. Alert on animal activity only. Best for trail cams and nature monitoring.
person car truck cat dog bear
Custom filters
Don't want a preset? Set any combination of allowed and ignored classes per camera. The AI Suggest feature analyzes your last 72 hours of detections and recommends which classes to filter based on what's actually showing up on each camera.
GPU Compatibility
Works with what you already have.
A graphics card is optional. An unmeasured card does not raise the camera count. Size the computer on the sizing page.
No card required
Detection runs on the CPU while something is moving. Recording does not wait for it. A normal desktop, including a Dell an office already owns, is a normal recorder.
Windows, when a card is present
Windows uses DirectML. The card we probed is an RTX 5060 Ti. Its decode did not raise the camera count on that 8-core desktop, because the CPU budget was already higher. Detection speed on a card has not been measured.
Linux
A machine with no NVIDIA card uses the smaller Docker image, tag edge. The gpu image is the large one, and only with an NVIDIA card, driver 580 or newer, and runs detection and video decoding on the card.
Smart Motion Zones
Draw zones. Eliminate noise.

Draw the driveway as a zone you care about, and the road as a zone you do not. A class filter can keep one camera to people, or to vehicles. Detection names what it sees. It does not decide whether the recording exists.

  • Per-camera zone drawing in the UI
  • Inclusion and exclusion zones
  • Combine with AI class filters (person only, vehicle only)
  • Time-of-day scheduling per zone
DRIVEWAY - INCLUSION
ROAD - EXCLUDE
GarageFront · 02/28/26 14:32
Smart Alerting
Alerts that learn what's normal.
LookoutNVR doesn't just detect objects - it classifies how unusual each detection is based on 7 days of history per camera. A person at the front door at 3 AM is treated differently than at 3 PM.
How the detection pipeline works
1
Motion Pre-Filter
Before running AI inference, each frame is checked for motion. If nothing has changed (below the per-camera motion threshold), the frame is skipped entirely. This saves GPU cycles on static scenes.
2
Detection
Frames with motion are named on the CPU, or on a graphics card when one is present. A detection carries a box and a confidence. The default confidence is 50 percent. The recording already exists.
3
Zone & Class Filtering
Detections are filtered against your inclusion/exclusion zones and per-camera class filters. A person detected outside your inclusion zone is discarded. A bird on a camera with the Farm preset is ignored.
4
Anomaly Classification
Each surviving detection is classified as Routine, Unusual, or Rare based on 7-day per-camera history. This determines how aggressively duplicate events are suppressed.
5
Alert Rule Matching
Detection events are checked against your alert rules. Rules can target specific cameras or be global, filter by object class and minimum confidence, and have independent cooldown timers.
6
Notification Dispatch
When a rule matches, notifications are sent via your configured channels - ntfy.sh push notifications, email, webhook, and/or desktop app toasts. All channels fire in parallel.
Anomaly Classification
Not all detections are equal.
LookoutNVR maintains a rolling 7-day frequency baseline for every object class on every camera. Detections are automatically classified into three anomaly levels that control event suppression and alert urgency.
Routine
≥20% of detections in last 7 days
Common activity for this camera. Events are suppressed for at least 5 minutes if the object hasn't moved. Your front door camera seeing a car every 10 minutes won't fill your event log.
Cooldown: 5+ minutes
Unusual
5-20% of detections in last 7 days
Uncommon for this camera. Events are logged with the standard cooldown. A dog at your garage camera that only shows up a few times a day.
Cooldown: 60 seconds (configurable)
Rare
<5% of detections or never seen
Very uncommon or first-time detection for this camera. No suppression - every detection is logged immediately. A person at a back-fence camera at 2 AM.
No cooldown - always passes through
Spatial awareness
Cooldowns only suppress stationary objects. If an object moves more than 10% of the frame width or height, the cooldown resets and a new event is created. A parked car is suppressed after the first event - a car driving through creates events as it moves.
Notification Channels
Get alerted your way.
Configure one or all channels. They fire in parallel - no delay, no single point of failure.
ntfy.sh Push
Push notifications to your phone via ntfy.sh - a free, open-source service. Subscribe to your topic in the ntfy app (iOS & Android) and get instant alerts with priority levels.
Works with self-hosted ntfy servers too.
Email (SMTP)
Send alert emails via any SMTP provider - Gmail, Outlook, Mailgun, your own server. Includes detection details, camera name, confidence score, and timestamp.
TLS on port 587. Configure host, credentials, and recipients.
Webhook
HTTP POST with JSON payload to any URL. Connect to Home Assistant, Node-RED, Slack, Discord, PagerDuty, or any automation you want.
{ event_type, camera, subject, message, timestamp }
Desktop App Toast
Native Windows toast notifications via the LookoutNVR desktop app. Click a toast to bring the app to the foreground and jump to the camera. Works via SignalR - real-time, no polling.
Also shows camera-offline and camera-online status changes.
Alert Rules
Fine-grained control over
what triggers a notification.
Alert rules sit on top of the detection pipeline. A detection event can exist in the timeline without triggering an alert - rules let you decide which detections are worth a push notification at 2 AM and which can wait until you check the event log in the morning.
Rule Setting What It Does
Camera Target a specific camera, or leave blank for a global rule that applies to all cameras.
Object Classes Which detection classes trigger this rule. e.g., only person and car. Leave empty to match all classes.
Min Confidence Minimum confidence score (0-100%) to trigger. Higher = fewer false positives, lower = more sensitive.
Cooldown Minimum seconds between triggers for this rule. Prevents notification storms. Default: 30 seconds.
Notification Channel Where to send: in-app, email, webhook, or push. Different rules can use different channels.
Enabled Toggle rules on/off without deleting them. Useful for temporary rules during construction, events, etc.
Example setup
Rule 1: "Night Watch" - All cameras, person only, confidence ≥70%, ntfy push, cooldown 60s, active 10PM-6AM.
Rule 2: "Driveway Arrivals" - Driveway camera, car + truck, confidence ≥50%, webhook to Home Assistant, cooldown 5 min.
Rule 3: "Package Watch" - Front door camera, all classes, confidence ≥60%, email notification, cooldown 30s, enabled only during delivery hours.
Notification Content
What you get when an alert fires.
NTFY.SH PUSH NOTIFICATION
person detected on FrontDoor
Rule 'Night Watch' triggered. person detected on 'FrontDoor' with 92% confidence.
Priority: high Tags: eyes 03/25/26 02:14:33
WEBHOOK JSON PAYLOAD
{
  "event_type": "detection_alert",
  "camera": "FrontDoor",
  "subject": "person detected on FrontDoor",
  "message": "Rule 'Night Watch' triggered.
    person detected with 92% confidence.",
  "timestamp": "2026-03-25T02:14:33Z"
}
Camera-offline and camera-online events also send notifications - so you know immediately if a camera drops, not when you happen to check the live view.
Built-In Testing
Test before you trust.
The settings page has a "Send Test Notification" button that fires a test alert through all configured channels. You'll see exactly which channels worked and which need attention - before you rely on them at 2 AM.
False Positive Reduction
Six layers between a pixel change
and your phone buzzing.
Most NVR software gives you motion detection and hopes for the best. LookoutNVR stacks multiple independent filtering stages so that by the time an alert reaches you, it's almost certainly real.
1
Motion Gate
AI inference is expensive. If nothing in the frame has changed, the GPU doesn't run. Static scenes are skipped entirely - no wasted compute, no phantom detections from sensor noise.
2
Confidence Threshold
Low-confidence detections are discarded before they enter the pipeline. The threshold is tunable - raise it on cameras with complex scenes, lower it on clean backgrounds where you need maximum sensitivity.
3
Duplicate Suppression
When the model detects the same object multiple times in overlapping bounding boxes, only the highest-confidence detection survives. One person = one detection, not three stacked on top of each other.
4
Spatial Zones
Draw inclusion and exclusion zones directly on each camera's view. A car on the public road doesn't trigger - but the same car in your driveway does. Per-zone class filtering narrows it further.
5
Class Filtering
Per-camera whitelist and blacklist of object classes. Your barn camera can ignore birds and cattle. Your front door camera can only care about people. Different cameras, different rules.
6
Anomaly-Aware Cooldowns
A stationary car in a parking lot doesn't generate an event every 5 seconds. The system tracks object positions and suppresses repeated detections of the same object in the same spot - but resets instantly when something moves or something rare appears.
The result

Zones tell a camera which part of the picture matters. Detection still only names what moved. A windy tree line costs more CPU than an empty hallway, which is why the sizing page asks about busy scenes.

Coming Soon
AI is just getting started.
Everything above runs locally on your hardware today. We're working on what comes next - cloud-connected features that build on your local detection engine without replacing it.
License Plate Recognition
Read plates from driveway and parking cameras. Allowlist known vehicles, alert on unknowns.
Audio Anomaly Detection
Detect glass breaking, alarms, shouting, and gunshots from camera audio streams. Per-camera baselines learn what's normal.
Cloud-Assisted Rules
Auto-generate alert rules and filter presets based on detection patterns across your cameras. Your data stays local - the cloud just helps tune.
Your cameras are working.
Is your NVR keeping up?
Download LookoutNVR free. AI detection, smart alerts, and per-camera filters - no subscription needed to get started.
Download Free for Windows See Pricing