Detecção de fogo baseada em sinais de vídeo por meio de análise de séries temporais

Abstract

In recent years, large forest fires have often been reported worldwide. When they are not controlled, they can have irreparable impacts such as loss of biodiversity, human health losses, rural properties, industries and all the communities that permeate forests. In addition, due to technological development and the need for higher areas of food planting, forests worldwide have been increasingly suppressed. In this context, it is essential to add effective techniques capable of automatically detecting fires. In this sense, in recent years, it is possible to observe a significant increase in the techniques that investigate automatic fire detection. These techniques are usually based on computational vision algorithms and can understand from video processing or images to the use of deep learning, especially convolutionary neural networks. This work is presented a method for detecting fire in two steps based on video signals by analyzing the temporal series. In the first stage, regions with the possibility of fire are selected through a spatial detector implemented by a convolutionary neural network. In the second stage, a temporal detector is used based on the analysis of the temporal series (AST) of the regions of interest identified by the space detector. AST consists of a sequence of steps: initially, the investigated video ROI is divided into blocks of dimension $ n \times N $ pixels and, for each of these blocks, an average temporal series is calculated. To improve the glow effect of fire, the first temporal difference is calculated, labeled and organized into input vectors to be classified by six machine learning algorithms. The values different from the size of the Block $ N \times N $ are analyzed, observing computational performance. In all 87 videos are used to evaluate AST, with 35 composing the training set of machine learning algorithms and 52 make up the test set. The AST approach was applied to ROI's with fire possibilities, identified by the space detector. The test set consists of 50 \% of the videos being false positive and 50 \% being true positive. Experimental results show that the addition of temporal information to the spatial detector produces significant error reductions. This reduction depends on the size of the block used. The best performance was obtained with regions of interest divided into $8\times 8$ pixels and SVM classifier. In this scenario, AST is able to eliminate all false positives without reducing the real positive rate by adding an average video processing time less than 1 second.