Step 01
Ingest and batch
Up to five RTSP cameras are decoded and batched together, so the GPU processes all of them in one pass.
All case studies
Case Study · Workplace · Security
Office Security System
Five ordinary IP cameras and one NVIDIA Jetson running DeepStream watch our own office around the clock. They report the shutter, the glass door, visitors, after-hours movement, lights left on and cameras that go dark, and post each alert with a photo to the team’s Google Chat.
At a glance
A night guard that never looks away
Live at our own office since August 2026
Input
Ordinary IP cameras at the entrance and inside the office
Output
Google Chat and Gmail alerts with photos, recorded clips, a nightly report and a chat bot
Approach
NVIDIA DeepStream on a Jetson, with an object detection model trained on the site’s own objects
Runs on
One device on the office network, with no video sent to the cloud
01 / The challenge
Cameras record. Nobody watches.
An office camera system records everything and tells you nothing. Problems surface the next morning, if anyone thinks to scroll back through the footage.
01
Was the shutter closed?
The last person out pulls down the shutter. Nobody else knows whether it went all the way down, or when.
02
The door left open
The glass door gets propped open for a delivery and stays that way through the afternoon.
03
Lights on all night
The floor empties at 7 PM and the lights, fans and AC keep running until the morning.
04
False alarms kill trust
A one-frame detection glitch, a passer-by or a stray dog must not ping the whole team.
The same entrance camera by day and in infrared at night · drag to compare
02 / How it works
From five camera feeds to one alert
One DeepStream pipeline on the Jetson. Video frames stay in GPU memory from decode to detection.
Step 02
Detect
A custom object detection model, optimised for the device, finds people plus rolling shutter, open/closed glass door and food container.
Step 03
Track and filter
An IOU tracker gives each person an id. A native C++ probe drops cars and animals and ignores single-frame “ghost” people.
Step 04
Apply the rules
Zones, timers and time windows decide what counts. State is saved to disk, so a restart in the middle of a five-minute door timer does not lose it.
Step 05
Alert with proof
Each alert carries a photo from the annotated stream, usually alongside a short strip showing who was there. It is sent from a background thread so the video never stalls.
Step 06
Ask it anything
The Google Chat bot answers questions from the live system: shutter, door, lunch, cameras and today’s alerts.
From the Footage
Real clips from our entrance camera
Unstaged recordings from the entrance camera. Every face on this page is blurred. The boxes come from the system’s own model, and the red bar names the alert for that moment.
Annotated output
Shutter: half open → fully open
At 8:29 AM the shutter is raised. The system keeps sampling it and only commits to a new state once enough samples agree. The “Opened” alert then reached Chat, together with a strip of who was at the door while it went up.
Security Alert - U1 Office Ground Floor Opened
Food container + person = lunch is here
A container on its own is ignored, because a tiffin left on the bench should not fire. The alert only goes out when a person and a food container are seen close together for a few seconds, and only once a day. If nothing arrives by the cut-off time, a “Food not arrived” alert goes out instead.
Food Alert: Food Arrived at U1 Office
A closed laptop at someone’s side
A detector can’t see a closed laptop held flat against someone’s side. So for each finished entrance clip, a small vision-language model running locally is asked what each person is holding. In the first test, this walk-past came back as “Laptop” and crops of ordinary visitors raised nothing.
Security Alert - U1 Office Ground Floor Office Property Carried at Entrance (Laptop)
Someone at the shutter at 5:40 AM
Overnight, anyone who stays in view of the entrance camera raises an alert. The camera switches to infrared after dark, so the same model keeps working at night.
Security Alert - U1 Office Ground Floor Person Outside After Hours03 / Every alert
From the shutter to the lights
Each card below is a real alert the system posted to our Google Chat, shown with its subject line.
ShutterShutter opened or closed
Every change between fully open, partly open and fully closed is reported, with a strip showing who was there. On weekends, a shutter left half open raises a maintenance alert.
Security Alert - U1 Office Ground Floor Closed
Glass doorGlass door left open
A wall-clock timer starts the moment the door is seen open, so a blocked view or a flickering detection doesn’t reset it. After 5 minutes the alert goes out, with the frames of whoever opened the door.
Security Alert - U1 Office Ground Floor Glass Door is Opened
EntrancePerson at the entrance 5+ min
Someone sitting or waiting at the entrance for more than five minutes. The alert comes with a strip of four frames covering the whole visit.
Security Alert - U1 Office Ground Floor Person at Entrance 5+ min
After hoursPerson outside after hours
Overnight, anyone who stays at the entrance raises the alert. Inside the office, anyone seen during closed hours raises one too.
Security Alert - U1 Office Ground Floor Person Outside After Hours
PropertyOffice property carried out
Laptops, keyboards, monitors and cardboard boxes leaving through the entrance. Lunch boxes and tiffins are excluded.
Security Alert - U1 Office Ground Floor Office Property Carried at Entrance (Laptop)
FoodLunch arrived, or didn’t
One “arrived” message per day when the delivery reaches the door. If the delivery window closes with nothing seen, a “not arrived” alert is sent instead.
Food Alert: Food Arrived at U1 Office
SafetyLights left on in an empty office
Once nobody has been on the floor for a while, the inside cameras check whether the lights are still on, at lunch and after 9 PM. The message asks someone to switch off the lights, fans and AC.
Security Alert - U1 Office Ground Floor Lights Left ON
OfficeOffice empty for 5 minutes
A stopwatch for each zone counts the minutes with nobody present. Once everyone has gone for lunch or for the night, a reminder goes out to check the fans, lights and AC. It is sent once per window and won’t repeat until someone has come back and left again.
Security Alert - U1 Office Ground Floor Entire Room is Empty
Camera healthCamera offline, covered or moved
A camera that drops out is cut from the pipeline without stopping the others, and reconnected once it’s back. Offline emails repeat until it’s back. A covered or turned camera is caught by comparing its view with a daytime reference.
Clustrex Security Alert - Main Entrance Camera - Went Offline
Night report9 PM night report
Every night at 9 PM: shutter, door, inside cameras and lights in one message, with a photo from each camera. The subject line says plainly whether the office is still open.
Clustrex Night Report - U1 Office Ground Floor - OFFICE STILL OPENEdge, Not Cloud
One Jetson. Five Cameras. No Video Leaves the Building.
Detection, tracking and every rule run on the device on the office network. Only the alert text and its photo go out to Chat.
Ask the System
A security guard you can message
Mention the bot in the team’s Google Chat space and it answers from the live pipeline, with photos. It runs on a small Apps Script relay and needs no extra cloud billing.
What you can ask
statusShutter, door, food, cameras and inside lights, with photosshutterThe shutter right now and every change todaydoorThe glass door right now and who was at it lastlunchWhether the food has arrived todayphotoLive photos from every cameracamerasWhich cameras are onlinealertsToday’s alerts, in order
Built to Stay Up
A security system can’t take a night off
01
Cameras drop, the pipeline doesn’t
A camera that disconnects is cut out and the grid reflows. The other cameras keep detecting without a pause.
02
Hot reconnect
A regular network check sees when the camera is back, then plugs it into the running pipeline with recording intact.
03
State survives restarts
The door timer, shutter state and daily food flag live in a small on-disk store. A reboot carries on where it stopped.
04
No lost alerts
On shutdown it waits for pending Chat and email messages and finishes writing recordings, so clips aren’t corrupted.
05
Smart Record
When a person is seen, the camera’s own H.264 stream is saved straight to MP4, bypassing the GPU and costing almost no CPU.
06
Docker, restart on failure
Everything ships in one image that restarts on failure, and the device reboots itself on a schedule.
04 / Limits and Next Steps
What we’re still working on
This system runs every day at our own office. These are the gaps we know about.
Current limits
TodayNo hardware video encoder
The Orin Nano has no NVENC, so the annotated live stream is encoded on the CPU. That caps how many cameras one device can show live.
Property checks are new
Laptops are covered. Keyboards and cardboard boxes still need staged walk-pasts before they can be trusted.
Each site needs its zones
Shutter, door and floor zones are drawn once for each camera, and a camera that moves needs them redrawn.
Next steps
NextTrain “carried laptop” in
Entrance frames are already labelled, so the main detector can learn carried items and drop the second model.
A live dashboard
The day’s timeline of shutter, door, visitors and alerts on one page, next to the Chat bot.
More sites, same box
Retail shutters, warehouses and clinics follow the same pattern of entrance, shutter and after-hours rules.
WhatsApp alerts
Send the same alert with its photo to the owner’s phone, for sites that don’t use Google Chat.
See It On Your Cameras
Ready to put your cameras on watch?
See what Korv Glanz would watch at your site, using the cameras you already have.
- ✓Personalized 1-on-1 walkthrough
- ✓A review of footage from your own cameras
- ✓Technical integration assessment
Shutter. Door. After hours. Lights. Camera health.
Request a Live Demo
Let's look at your site.
Tell us a little about your premises and our team will get in touch.
