Research On Smoke Detection In Natural Scene Based On Video

Wildfire is a serious threat to ecological system and human lives.Fast wildfire detection and continuous surveillance are crucial for damage minimization.With the wide application of surveillance cameras in wildfire prevention,more and more video-based smoke detection algorithms have been studied.Existing video-based smoke detection algorithms have achieved good performance under certain conditions.There are few public wildfire videos,which have a low resolution of hundreds by hundreds.However,the surveillance video currently used to monitor the vegetation coverage area is primarily high resolution video.Therefore,few video-based smoke detection algorithms can be applied to long-term monitoring of vegetation coverage areas.Currently,the demand for smoke detection algorithms based on high-definition video in natural scenes is increasing.But the smoke video required for research is difficult to collect and the public high-definition video data is lack,which lead to less research on smoke detection algorithms based on high-definition video in natural scenes.In this paper,detecting smoke region in high-definition video in natural scenes is studied using traditional methods of classification,classification neural network and segmentation neural network.A smoke region segmentation data set is collected and manually labeled.The main works of this paper are summarized as follows:(1)In this paper,four motion detection algorithms are compared to detect smoke motion in video.HI is used to evaluate five common color spaces and an improved color space to distinguish smoke region from background.Two feature descriptors and classifiers are tested to classify the smoke images.The motion detection algorithm is used to obtain the candidate regions in the video sequence.Then,the candidate regions larger than 10 × 10 pixels are converted into HSV color space and the spatial texture feature of candidate regions are extracted.Finally,the PCA and classification are employed to the spatial texture feature of candidate region.In the experiment of this paper,the true positive rate of smoke images classification is 89.55%.However,the shape distortion and the small field of view of the candidate region cannot be improved using the propose method.Moreover,the classification accuracy is not high enough.Therefore,the proposed method is not sufficient for video-based smoke detection in natural scene.(2)In order to improve the low accuracy of smoke image classification based on traditional algorithms,classification neural network is used to classify smoke images for detecting smoke in video in natural scenes in this paper.Three classical classification neural networks and two improved classification neural networks are compared for smoke image classification.Compared with traditional methods,the accuracy of smoke image classification using neural network is much higher.Nevertheless,the smoke detection method using motion detection and classification neural network still cannot improve the shape distortion and the small field of view of the candidate region.(3)In order to avoid candidate region distortion and improve small field of view,a modified 3D fully convolutional neural network model and a 3D parallel fully convolutional neural network model are proposed to segment smoke region in the video sequence in this paper.Experiments show that the proposed model cannot segment the whole smoke region when the smoke is far,small,or slow moving.However,the proposed model is real-time and robust to clouds,fog,and many other interfering objects in natural scene.