The occurrence of forest fires not only causes serious damage to the forest ecosystem,but also poses a serious threat to the safety of public life and property.Therefore,the early prevention and control of forest fires has become the primary task of forest resources protection in the world.Video monitoring technology has been widely used in early forest fire prevention and control due to its wide coverage and fast response speed.The generation of smoke is an early sign of forest fire.The video information is collected in real time by the camera installed in the tower,and the smoke detection algorithm based on video and image can detect the forest fire in time.However,the traditional smoke feature extraction detection method has a relatively single application scenario and a high smoke detection rate.Since the diffusion movement of smoke is an important indicator for early forest fire smoke detection,this paper studies the dynamic spatio-temporal information in forest fire videos and designs a video smoke detection model based on deep learning,aiming to improve the accuracy of smoke detection and provide necessary technical support for forest fire prevention and control.Specific work contents are as follows:(1)The video image smoke detection data set and video sequence smoke detection data set are constructed.Firstly,considering the real scene environment of different wind speed,jitter of the pylon,illumination and other factors,the smoke and non-smoke videos of the real scene were collected through the webcam installed on the observation tower.Second,collect publicly available early smoke videos containing forest fire scenes.Finally,the data set was constructed according to the difference of smoke color and shape,and the annotation software Lableimg was used to annotate the two data sets.(2)A video image smoke detection method based on Faster R-CNN was studied.In order to make full use of the spatial location and semantic information of single frame smoke image,the multi-scale smoke image feature extraction and visual field expansion were carried out on the model respectively in view of the characteristics that the scale of smoke changes dramatically in the early stage of forest fire preheating.The smoke detection data set of video images including smoke and non-smoke constructed in this paper is used for verification.The experiment shows that the detection model can locate the location of smoke in forest fire videos and effectively improve the accuracy of smoke identification.(3)A smoke detection method based on global-local feature fusion is designed for video sequences.To make full use of the fire video temporal information,this paper takes Faster R-CNN video images of smoke detection model,the high-level semantic extracted from the candidate network in the smoke region were fused with associated modules,and long-term memory module is introduced to further increase the scope of feature fusion.The smoke detection data set of video sequence constructed in this paper is used for verification.The experiment shows that the global-local feature fusion module can effectively suppress the smoke false alarm rate and improve the smoke detection rate. |