Research On Forest Fire Smoke Detection Based On Spatial-Temporal Analysis Of Video Images

Due to global climate changes and human activities,it is urgent for us to strengthen prevention of forest fire.Among many forest fire monitoring methods,video surveillance is widely used in forest fire detection for its strong visibility,high flexibility and low cost.As the major manifestation of the forest fire at the early stage,the smoke can play an important role in the detection and location of forest.Most existing methods of forest fire smoke detection based on video are based on pixel change analysis,through which smoke features are extracted and detected.However,in the early stage of smoke or when the position of smoke is far from the camera,the smoke on the video image is small and it’s spreading is irregular.Together with the changing background environment,smoke features based on pixels are not very obvious.Therefore,it is more difficult for us to rely on pixels to automatically identify the smoke region in the image.A set of forest fire smoke detection and identification methods have been established based on image block,which combine the digital elevation model and video pictures in light of principle of visible video image processing.Firstly,the image matching technique and atmospheric scattering model is taken advantage of to realize image enhancement,which will weaken the impact of the distant on image.Then we use the space-time features of regions with smoke to identify suspected smoke blocks in the image.Later color and texture features of smoke is combined to build classification model based on random forest model,realizing the identification of smoke regions and non-smoke regions.In this study,the region of Jiulongshan Nature Reserve in Mentougou District,Beijing is selected as test region,and have verified the algorithms of the smoke detection based on video.The main research content of this article is as follows:1.The study of image enhancement technology based on spatial distance.We have used ray tracing algorithms to produce the oblique view of DEM,also called depth image in which pixel value means spatial distance from the observation point to DEM.Then the depth image and the images from the video are matched by topographic feature points.The atmospheric scattering model based on the depth image is used to enhance the live pictures and partition adaptive regions,reducing the ray attenuation caused by atmospheric scattering model.It will eliminate the effect of distance on the pixel intensity to restore the clarity of the image,which plays a positive role in the detection and identification of smoke in the distance.2.The extraction of the suspected smoke block based on regional dynamic characteristics.The video will be transformed into continuous image sequence which are produced at the rate of one frame per second.Then the regional image brightness of the continuous image sequence is used to calculate the signal-to-noise ratio of the image sequence among different regions,from which the adaptive threshold range is determined.The time window is set up to analyze image blocks if the signal to noise ratio exceed the threshold value.Based on regional dynamic characteristics,the detection method of suspected smoke is proposed to realize the image regions in the complex environment.3.The classification of suspected smoke based on machine learning.The smoke and non-smoke samples are used to build support vector machine model and random forest classification model.The LBP texture and colour features of the suspected smoke blocks is selected as the main features.The classification results are compared and the classification based on random forest model is the most accurate.The study of suspected smoke classification based on machine learning has improved the detection accuracy,which has reached 93.26%.The innovations in this research is as follows:(1)The image enhancement and partition technology based on the depth image.The depth image is created from DEM and is used to establish image enhancement model and image partitioning,and the model and algorithm based on distance is realized to image adaptive partition and enhancement.(2)Smoke detection technology based on regional dynamic characteristics.Based on the signal-to-noise ratio algorithm of image sequence,the brightness of the local area is counted,then the multi-scale spatial-temporal features of the image sequence are extracted,and the position and time are determined which is affacted by smoke occurance.Finally,the non-smoke regions is removed based on machine learning.