Awareness at the edge.
Smart security camera
On-device person detection makes camera alerts more useful while reducing video uploads.
- Industry
- Commercial security
- Our role
- Detection & edge deployment
- Core tools
- Jetson Nano · TensorRT · OpenCV
A practical problem.
Motion-only alerts created too many false alarms. Cloud video analysis added bandwidth costs and delayed notifications, so the client needed detection that could run on affordable local hardware.
- Engagement
- 12 months · Q1 2023–Q1 2024
- Project setting
- Commercial security integrator · UK
Built around the constraints.
- Recognise people.
A lightweight detector is trained for the client’s indoor and outdoor camera views.
- Optimise for the device.
Quantisation and TensorRT reduce inference time on Jetson Nano.
- Send useful alerts.
Local processing sends event metadata and alert thumbnails instead of continuous video uploads.
How it comes together.
Video stays close to the camera: local detection and tracking turn frames into alerts for the existing security system.
-
Capture
RTSP streams · USB cameras
The capture module accepts camera feeds and negotiates resolution and frame rate for processing.
OutputFrames for local inference
-
Detect
YOLOv7-tiny · TensorRT · Jetson Nano
An optimised person detector runs on-device. Quantisation and layer fusion reduce the inference workload.
OutputPerson detections & confidence
-
Track
DeepSORT · Multi-object tracking
Detections are linked across frames to maintain individual tracks as people move through the camera view.
OutputTracked people across frames
-
Notify
VMS integration · Alarm panels
Zone-specific thresholds determine alerts. The system sends event metadata and thumbnails to existing monitoring tools.
OutputAlerts, event metadata & thumbnails
Results from this project.
- Person detection
- 99.2%Measured in daytime conditions
- Typical inference
- 30 msDown from 45 ms on Jetson Nano
- Less bandwidth
- 75%Compared with continuous cloud upload
Before
Motion triggers produced frequent false alarms; cloud analysis required continuous video uploads.
After
Local person detection sends focused alerts, event metadata, and thumbnails to the existing security system.
In context. Detection accuracy falls to 96.8% in nighttime infrared mode. Crowded scenes can increase inference time to approximately 45 ms.
Working on something similar?
Tell us about your system and where you need support.