Design And Implementation Of Fire And Smoke Detection System Based On Improved SSD

Traditional feature extraction algorithms are used for fire and smoke detection.It is difficult to extract all fire and smoke features and high-level feature information of pictures.Convolutional neural networks have good performance in feature extraction capabilities in the image domain.This paper researches fire and smoke detection by using Single Shot Multi Box Detector(SSD),which has a good balance between accuracy and operating speed and it can meet the accuracy and real-time requirements of fire and smoke detection.The main work of this paper is organized as follows:(1)Improve SSD model.First,making the SSD more lightweight by using Mobile Net V2 as the backbone network of the SSD and replacing the standard convolution in the detection branch with a deep separable convolution;replacing the standard convolution in the extra feature extraction layer of the SSD with an inverted residual Difference module to improve the feature extraction ability of the model.Second,improving the detection branch of the SSD to a Feature Pyramid Network(FPN)structure to solve the problem of non-circulation of different detection branch information of the SSD,and enriching the input feature map information of the detection branch by fusing the information of different levels of feature maps.Third,replacing the additional feature extraction layer of SSD and the standard convolution in the detection branch with deformable convolution to improve the geometric modeling capabilities of the model.Fourth,the attention mechanism is introduced into the SSD,so that the network can efficiently focus on the characteristics of the fire area.(2)An improved multi-task network is proposed.Aiming at the problem of inconspicuous smoke boundaries and weak smoke characteristics which are lead to poor smoke detection effects,smoke detection tasks are replaced with smoke recognition tasks.A multi-task network is improved on the basis of SSD.The model includes smoke recognition branch and fire detection branch,through sharing The bottom features of the smoke recognition branch and the fire detection branch realize different stages of fire detection,simplifying the structure of the model and enhancing the detection ability of the model.In order to reduce the model parameters,this paper uses the idea of decomposing convolution to improve a lightweight convolution on the basis of deep separable convolution for multi-task network.(3)A fire detection system is realized.The hardware and software of the system are designed.With Jetson Nano as the hardware platform,the hardware system is integrated,and a fire detection software system with a visual interface is developed.The picture information is captured from the camera and the detection result is displayed on the interface.In this paper,the improved algorithm is verified by experiments,the detection performance is tested on the fire test set,and the reasoning speed is tested on the Jetson Nano development board.The improved fire detection method in this paper is effective in terms of detection rate and real-time performance.