Design of Machine Learning-Based Smoke Surveillance

A real-time machine learning-based fire smoke detection method that can be incorporated with a automatic monitoring system for early alerts is proposed by this paper. The successive processing steps of our real-time algorithm are using the motion segmentation algorithm to register the possible smoke position in a video and then analyze the spectral, spatial and motion orientation characteristics of the smoke regions in the image sequences. Characterization of smoke was carried out by calculating arithmetic mean and standard deviation from the extracted feature vectors, and the non-linear classification method using support vector machine is applied to give the potential fire smoke candidate region. Then, the continuously adaptive mean shift (CAMSHIFT) vision tracking algorithm is employed to provide feedback of the fire smoke real-time position at a high frame rate. Experimental results in a variety of conditions show the proposed support vector machine-based fire smoke detection method is capable of detecting fire smoke reliably and robustly.