Fire Detection Using Dense Trajectories

Abstract

This paper proposes an automatic computer vision-based system for fire detection in videos. There are many previous methods for video-based fire detection but only very few of them have considered the challenge of camera motion or motion of the background scene while finding features based on motion of fire. Our method is divided into two phases. First, we train our system for color characteristics of fire with the help of Gaussian mixture model (GMM), and for texture features which are computed using local binary patterns (LBPs). Next, dense trajectories are computed for motion features which are free from camera motion or challenges of moving scene. Bounding boxes are detected with the help of color and texture models. Subsequently, dense trajectories are projected onto codebooks for feature vector computation, and chi-square kernel-based SVM is employed for classification of fire and non-fire motion representations. Quantitative evaluation of our method indicates the fitness of temporal features for fire detection.