Research On Video Smoke Detection Based On Spatial-temporal Features

Fire is a threat to the lives,health and property safety of humans and animals.Fire early warning is a powerful means to combat fire and minimize losses.The detection of early smoke before the fire is an important indicator of fire warning.Currently,many smoke detection methods have emerged to enable simple smoke detection and alarm.Besides,the researchers are focus on improving the accuracy,speed and adaptability of smoke detection.At present,video smoke detection technologies usually design the smoke features manually and then classify by the traditional machine learning algorithm,which extracts the spatial and temporal feature simply for comprehensive judgment and the results of these algorithms are easily affected by the environment.and the others are extracting features automatically by deep learning network,which is a new view of smoke detection.However,whether the detection algorithm is based on manual design features or deep learning,it is a challenge to fuse the spatial and temporal features and detect smoke quickly,accurately and economically in a complex environment.Based on the above issues,some research has been done as follow:At first,do some research on moving target segmentation,and propose an improved suspected smoke area segmentation method.Foreground extraction based on the difference between the wavelet low-frequency component of the background model established by the mixed Gaussian model algorithm and the original video frame,which improves the problem of the inner hole of the suspected smoke area.Secondly,due to the spatial and temporal characteristics of smoke,an end-to-end dualpath temporal and spatial smoke detection network was established: based on the Pseudo-3D Res Net network structure with attention mechanism,the temporal and spatial characteristics were fused to improve the smoke detection network's efficiency,and the network input is the fixed length continuous video frames.Thirdly,a framework of the smoke detection system has been proposed: block the frame and filter out the background block by background subtraction,establish a specific dataset and the whole smoke detection system.The network detection speed is about 18 FPS,and the accuracy rate is 98.36%.It performs good especially for the too thin and thick smoke,and achieves the real-time requirements.Compared with other algorithms,the detection method proposed in this paper based on the spatial-temporal characteristics of smoke has improved the detection accuracy and detection rate.Combined with deep learning algorithm,a more robust smoke detection method is proposed,which has certain practicality.