Researches On Smoke Detection Technology Based On Visual Attention And Spatial-temporal Characteristics

Fire is one of natural disasters which have serious threat to people’s lives andproperty safety. In recent years, fire frequently occur and have brought much loss topeople’s lives and property in many places such as large shopping malls，forest andhomes. Since smoke is an early product of fuel combustion, so it could be animportant object for early fire detections.As for some visual features of fire smoke on image frames, the smoke detectionalgorithm based on the visual attention and spatial-temporal characteristics is mainlyfocused in this paper. Firstly, significant figures are generated using the visualattention on smoke brightness characteristics to find the suspected area of the smoke.Then, extract the spatial-temporal features of smoke on suspected area. Finally, smokeis detected using the Support Vector Machine classifier (SVM). This research cansummarize in the following respects:1. After having a study on the previous research articles, smoke detectiontechnologies are summarized in this paper and difficulties in on-line smoke detectionsare analyzed, and the roadmap of the smoke detection is proposed foropen environments.2. The visual attention models are utilized in this paper to extract the region ofinterest of the smoke video image in open environment. Since the traditional visualattention model is implemented using the Gaussian pyramid, having a huge amount ofcomputation, so it’s not suitable for real-time detection of smoke. So in this paperFirstly, a visual attention model of smoke brightness is used to extract the suspectedareas of the first frame image, and the Kalman Filter Background Updating Model(KF) is then taken to track a series of smoke region frames, and ultimately thesuspected smoke areas on the entire video images are generated.3. Aiming at the dynamic texture features of the smoke suspected areas in videoimage, the smoke detection algorithm based on the Spatial-temporalLocal Binary Pattern (LBP) is proposed. In order to more accurately describe thetexture features of smokes, this paper introduce the dynamic texture features on threeorthogonal planes of XY, XT and YT.4. In this paper, some deep study is taken on pixel’s motion features in suspected smoke areas. The mutual information of smoke pixels’ movements in two-dimension,such as speed and direction of the suspected smoke area, accurately reflect motionrelevancy between two sequential smoke regions.5. In allusion to the contour volatility of suspected smoke areas，a new detectionalgorithm is proposed in this paper based on the gradient information entropy of thepixels’ tangent angles of suspected smoke contours in adjacent frames of videoimages.6. In this paper, the support vector machine classifier is designed to discriminateagainst smoke. Firstly, using a lot of smoke and non-smoking video images to trainthe classifier, and then using this classification to finally determinate smoke samplevideo images. After a great number of sample experiences, the proposed smokealgorithm has higher accuracy, and could satisfy the requirements of the earlydetection for forest fires.In conclusion, the smoke detection algorithm based on the visual attention andSpatial-temporal characteristics could achieve early fire detection and alarm functions,putting forward a novel solution of the intelligence and automation to early detectionfor forest fires.