Standart IP kamera
Özel bir kamera gerektirmez; yaygın kullanılan IP kameralarla test edildi.
An alarm the moment a dangerous object appears
A custom-trained YOLOv5 model processes security camera footage in real time; when a blade or weapon is shown, a siren sounds and the people in charge get SMS and e-mail alerts.
- Industry
- Security
- Completed
- 2024
- Technologies
- Yolo AINVIDIA CUDAPyTorchOpenCVPython
%92
Successful detection in lab testing
3
IP cameras at different points
3
Alarm channels: siren, SMS, e-mail
The problem: the eye on the screen cannot catch everything
At security-sensitive points cameras record all the time, but an operator watching live is looking at many screens at once. When attention drifts or the operator is away from the desk, a critical moment is easily missed; the footage is only reviewed after the event.
The aim of this R&D work was to have a machine watch the footage too, as an aid to the human eye:
- Detect a blade or weapon the moment it appears on camera
- Deliver the alarm both on site (siren) and to remote staff (SMS, e-mail)
- Work with widely used standard IP cameras
Nasıl çalışır
Three steps from camera to alarm
The camera footage is fed to the model frame by frame; when the model finds a dangerous object, the alarm channels are triggered at once.
-
Footage 3 IP cameras
-
IP camera Hikvision DS-2CD2T63G2-2I at different points
-
Video stream Frames are captured continuously
Frames -
-
Inference GPU server
-
YOLOv5 model Custom-trained; blades and weapons
-
Confidence threshold Low-probability detections are dropped
Detection -
-
Alarm Instantly
-
Siren Audible warning on site
-
SMS and e-mail Notifications to the people in charge
-
The model
Not a general model, one trained for this job
Off-the-shelf object recognition models are trained for everyday objects; they are not enough to tell blades and weapons apart reliably. So YOLOv5 was trained with data prepared specifically for these objects.
YOLOv5 processes an image in a single pass and gives both what an object is and where it is in the frame, which suits real-time processing of a live camera stream.
Test ortamı
Laboratory test, July 2024
| Component | Used |
|---|---|
| Software | |
| Model | YOLOv5, custom-trained |
| Detected objects | Blades and weapons |
| Hardware | |
| Camera | 3 × Hikvision DS-2CD2T63G2-2I IP cameras at different points |
| Processing | NVIDIA GeForce GTX 1080 Ti |
| Alarm | |
| Channels | Siren, SMS, e-mail |
| Result | |
| Successful detection | 92% (laboratory) |
The result was measured in a laboratory. In the field, lighting, angle, distance and crowding can affect it.
Example scenarios
Where would it help?
Building entrance: the moment a knife appears in a visitor's hand on the entrance camera, a siren sounds at the security point and the head of security gets an SMS.
Waiting and service areas: in a crowded hall, even while the operator is watching other screens, a blade being shown is reported by e-mail and SMS.
Nights and off-shift hours: the alarm reaches the people in charge even when nobody is at the monitoring desk.
These are example uses the system is aimed at; the work was tested in a laboratory.
Design highlights
Teşhir anında algılama
Nesne görüntüde belirdiği anda işlenir; kayıt sonradan incelenmeyi beklemez.
Çok kanallı alarm
Siren sahadakileri, SMS ve e-posta uzaktaki sorumluları uyarır.
Özel eğitilmiş model
Model bu nesne sınıfları için hazırlanan veriyle eğitildi.
What to consider when moving to the field
Although the laboratory result is promising, a real security point brings different conditions. A field deployment also covers:
- Measuring again under the lighting, angle and distance of the installation site
- Tracking false alarms and tuning the alarm threshold to that measurement
- Linking the alarm to the existing security procedure and video management system
- Processing footage in line with personal data law (KVKK)
Technology
Technologies
AI / ML
Yolo AI
NVIDIA CUDA
PyTorch
Backend
OpenCV
Python
Considering image processing for your own site?
Let's plan it together, from model training to camera and alarm integration.
Request a quote
Call us+90 850 840 25 29
Have a similar project?
Share your needs and we will design and deliver a tailored solution.