Production · Computer vision
Real-time incident detection on live video
~30 fps camera streams classified into 5 incident types, with confirmation logic so a single ambiguous frame never raises an alarm.
- ~30 fpslive streams
- 5incident types
- 3 sconfirmation window
Architecture
- 01Cameralive ~30 fps
- 02OpenCVframe processing
- 03Classifierfine-tuned model
- 04Confirm3 s + score
- 05Alertoperator notified
Context
Nextronic builds smart video-surveillance systems. Operators can’t watch every camera, so the system has to watch for them — and only speak up when it matters.
The challenge
Classify incidents such as fire or intrusion on live streams fast enough to be useful, without drowning operators in false alerts triggered by one ambiguous frame.
What I built
- A real-time pipeline reading live camera streams at ~30 fps with OpenCV.
- A training dataset assembled from Kaggle and targeted scraping for each incident type.
- A pre-trained Hugging Face model fine-tuned to classify 5 incident types.
- Confirmation logic: an alert fires only when a detection persists for 3 seconds and passes a confidence score.
Outcomes
- 01Live classification of 5 incident types.
- 02Alerts on confirmed detections only, not single-frame noise.
- 03A dataset pipeline that extends to new incident types.
Stack
- Python
- OpenCV
- TensorFlow
- Hugging Face
- NumPy
Facing a similar problem?
I’m happy to talk it through — whether you’re hiring or need a freelance mission.