Spatio-temporal Deep Neural Network Based Video Smoke Detection

Smoke detection is the focus of fire detection technology,because it plays a key role in the early stages of the fire.There are many problems in the traditional sensor smoke detection system.Video smoke detection solves lots of problems brought by traditional smoke detection.Currently,most of the video smoke detection systems are relatively simple and have a high error rate.Therefore,this thesis combines some video smoke detection and video behavior detection technology,then builds a video smoke detection system based on spatio-temporal neural network.The video smoke detection system processes multiple frames of images,so removing most of the non-smoking regions can greatly reduce the time complexity of the system.Because of the uncertainty of motion direction and irregular shape of smoke,this thesis uses block-based motion detection to detect the moving region quickly and effectively.Then,in order to further reduce the detection area,dark channel priori smoke detection is used to exclude some non-smoking areas.After filtering out most of the non-smoking regions,the temporal and spatial features of video are extracted for classification and recognition.By combining 3D convolution neural network with DenseNet,this thesis proposes a kind of neural network for smoke video processing.At the same time,in order to reduce the model parameters,3 × 1 × 1 and 1 × 3 × 3 convolution kernels are used instead of the original 3 × 3 × 3 convolutional kernels.Finally,Global Average Pooling is extended to the time dimension to replace the original fully connected layer of 3D convolution neural network,which further reduces the model parameters.Through the above operation,we build a spatio-temporal neural network to extract the spatiotemporal features of smoke video.In this thesis,we use Tensorflow,a popular machine learning framework,and use Python to implement the whole video smoke detection system.At the same time,the performance of the algorithm is verified on the video smoke detection data set.And compared with the classical video smoke detection algorithm,the experimental results show that the method proposed in this thesis improves the accuracy and decreases the false detection rate.