Smoke Detection Based On Video

Recently, major fire accidents frequently, which caused huge losses to people's lives and property. However, the traditional fire detector itself has some shortcomings,such as the limitations of detection scene and vulnerable to the distance from fire source to detector devices, so it can not satisfy the current demand for fire alarm. Fire alarm technology based on video has become hotspot in recent years, because its advantage of wide application scenarios, simple testing equipment, high sensitivity and etc.In this paper, we study the smoke detection algorithm based on the video. The algorithm include smoke candidate region extraction, feature extraction and discriminant analysis. Firstly, we extract the smoke candidate region using adaptive Gaussian model, and color and texture features filter algorithm. Then we extract features of Gabor statistical texture, the rate of change of area and circularity, and direction of the main movement. Finally, we use SVM and threshold discrimination law, determine whether there is smoke appears. Meanwhile, we increase the speed of the algorithm using the principal component analysis(PCA) algorithm to reduce the dimensions of feature vector. We optimize the parameters of the SVM algorithm using PSO algorithm that the evaluation function is the average accuracy of the cross-validation algorithm, and the range of particle swarm initialization is region obtained from grid-search algorithm.In this paper, the test results of the test set with 1000 positive samples and 1000 negative samples, is detection rate of 92% and false positive rate of 16.6%, using SVM algorithm and Gabor statistical texture feature. The detection rate is 87.78% and the false positive rate is 0.79% for the test set of 9 videos, when we add smoke candidate region selection algorithm and dynamic features. The algorithm of this paper is applicable to the complex scenes which include the movement of vehicles and pedestrians.