Optical Flow Feature Based for Fire Detection on Video Data

Abstract

A fire detection on video data using optical flow feature is presented to improve the performance of detection when using only texture or color feature. We compare two kinds of optical flow that are dense optical flow using Farneback algorithm and sparse optical flow using the Lucas Kanade algorithm. The fusion of optical flow feature and Local Binary Pattern (LBP) as a texture feature is used to classify the video frame as fire or not fire using Support Vector Machine (SVM). There are three phases for fire detection in our framework. First, segmentation on each video frames based on Hue, Saturation, Value (HSV) color space is done to obtain the candidate of the fire area. Second, feature extraction using optical flow and LBP method is done to achieve the movement and texture features of the fire. Finally, the extracted features are classified to fire or not fire using the SVM method. The model is evaluated using stratified 10-folds cross-validation to be separated into learning process data and validation data. The best result is obtained using the Lucas Kanade optical flow feature and using a linear kernel SVM with 96.21% in accuracy.