Video Smoke Detection Classifies Moving Texture Rather Than Particles
Video smoke detection turns an ordinary camera into a fire sensor by analysing the image instead of sampling the air. Smoke has a distinctive signature in video: it moves upward and spreads, it lowers local contrast and blurs edges behind it, its colour is desaturated, and its motion is slow and turbulent rather than rigid. Classical implementations combined background subtraction with these motion and texture cues; current systems train a convolutional network on labelled smoke imagery. The reason it exists is geometry rather than sensitivity: in a large atrium, an aircraft hangar, a warehouse or a tunnel, smoke may never reach a ceiling detector in a usable time, but it is visible from the moment it forms. The costs are equally real. It sees only what is in frame and in focus; it is fooled by steam, dust plumes, headlights, moving shadows and fog; and its performance is a property of the training data as much as of the optics. After this Concept you can say what visual evidence a smoke classifier is actually using, and name the scenes that will fool it.
This Concept is waiting for its first lesson!
Video smoke detection turns an ordinary camera into a fire sensor by analysing the image instead of sampling the air. Smoke has a distinctive signature in video: it moves upward and spreads, it lowers local contrast and blurs edges behind it, its colour is desaturated, and its motion is slow and turbulent rather than rigid. Classical implementations combined background subtraction with these motion and texture cues; current systems train a convolutional network on labelled smoke imagery. The reason it exists is geometry rather than sensitivity: in a large atrium, an aircraft hangar, a warehouse or a tunnel, smoke may never reach a ceiling detector in a usable time, but it is visible from the moment it forms. The costs are equally real. It sees only what is in frame and in focus; it is fooled by steam, dust plumes, headlights, moving shadows and fog; and its performance is a property of the training data as much as of the optics. After this Concept you can say what visual evidence a smoke classifier is actually using, and name the scenes that will fool it.
Are you a teacher? Sign in to start contributing.
Sign In