Real-time smoke detection system

About the client

An Austrian-based manufacturing company with a strong focus on process safety.

Mission behind the project

The client asked XPG Factor to create a next-level solution that offers more than traditional video cameras, which only capture video without analyzing it. Our goal was to develop a system that could recognize smoke in the frame as early as possible, in real time, while avoiding false alarms from moving objects, light, etc.

Issues we had to solve

One of the main challenges in developing the solution was smoke detection under non-ideal conditions, such as unstable lighting, variable camera angles and differing distances. All of these factors resulted in dust, shadows or glare. To minimize the number of false alarms, we put certain parameters into the algorithm that analyzed the object’s transparency level, motion features and other characteristics.

Implemented solution

As a result, our team has developed an intelligent software algorithm capable of analyzing video streams from a standard camera in real time and detecting even barely visible smoke. The solution requires no additional sensors and works stably in visual noise and complex lighting conditions.

Key features of the system

• Moving object detection
  The algorithm identifies active zones in the frame by filtering out static background and irrelevant elements.
• Visual feature analysis
  It evaluates the color, transparency and texture characteristic of smoke.
• Trajectory and motion pattern assessment
  The solution takes into account the smooth vertical motion of smoke, which is different from the movements of people, machinery and other mobile objects.
• Contextual frame sequence analysis
  The system compares current and previous frames to eliminate false alarms caused by glare and other background fluctuations.

Testing

Our team tested the system under a variety of conditions to maximize its reliability and accuracy. We used a wide range of video footage, from clean and well-lit shots to complex scenes with lots of visual clutter. In addition, our team simulated scenarios with multiple smoke-like objects to ensure that the system does not react to false signals. As a result, the algorithm showed stable operation and high recognition accuracy in all tested cases.