Real-time fire and hazard detection from video
Live and recorded camera feeds are analyzed in real time, with fire and flame events flagged and streamed to a browser dashboard.
The challenge
Monitoring teams couldn't watch every camera continuously, so fire and flame events risked being noticed too late.
What we built
- Video frames annotated with bounding boxes for fire and flame classes and exported in YOLO format.
- A custom-trained YOLOv8 model tuned to a 0.7 confidence threshold to keep false alarms low.
- A centroid tracker assigns unique IDs so each event is counted once.
- A Flask service streams the annotated video to any browser and exposes live counts through a JSON API.
Results
- Continuous automated monitoring of every feed
- Events counted once, not per frame
- Live dashboard and API for alerting systems
Step by step.
Ingest
Live or recorded feeds are read frame by frame.
Detect
A custom YOLOv8 model flags fire and flame.
Track
Unique IDs prevent double-counting of the same event.
Stream
Annotated video and live counts are shown in the browser.
More computer vision projects.
Automated visual inspection on a camshaft production line
A line-side camera reads each camshaft's engraved serial number and flags surface defects in real time, logging every part for traceability.
Highway traffic counting for outdoor advertising analytics
A camera on a highway billboard counts and classifies passing vehicles, producing impressions data advertisers use for media planning.
Let's see what this could look like for you.
Tell us about your process and systems. We'll outline an approach, the guardrails it needs and a pilot plan.
© Copyrights 2026 by Techno Genesis Software Solutions Pvt. Ltd.