Real-time Smoke Detection Based On Deep Learning

Thanks to the rapid development of artificial intelligence in the field of vision,the visual camera equipped with artificial intelligence model will play a more and more important role in all kinds of disaster early warning.Especially in various technologies of fire early warning,video smoke detection technology based on visual camera has attracted extensive attention because of its wide applicability,flexible deployment,rapid detection and high real-time performance.And there has been a trend to replace the traditional sensor-based early warning device.However,although video smoke detection has made considerable achievements,there are still many problems in related research.Firstly,the fire scenarios are very diverse,which leads to different characteristics such as the shape and movement of smoke generated at the fire scene,and the scenes of smoke data contained in the current public data set are very limited,which leads to the insufficient ability of smoke detection model to adapt to the real scene;Although many researches based on deep learning improve the accuracy or solve the problem of insufficient data set compared with other original methods,there is no research on improving the model accuracy from the perspective of designing network structure for the specific problem of smoke detection;And the huge model parameters of these smoke detection models make them unable to detect in real time,let alone deploy them to the edge with lower computing power.These problems restrict the application of smoke detection research to practice.Based on the above problems,this paper gives some improvements and feasible schemes respectively.The main work of this paper is as follows:1.Two smoke datasets are collected,which greatly supplement the existing public smoke datasets from multiple dimensions such as quantity,scene diversity,smoke characteristics and interference factors.This paper summarize the information about scene,smoke characteristics,interference factors and other dimensions for public datasets.Based on these information,the insufficient data quantity and diversity of these datasets in multiple dimensions are analyzed,such as the lack of smoke data in night scene,a large number of smoke data are not collected from the real environment,the background of smoke video is fixed,etc.Aiming at these shortcomings,this paper provides a set of targeted video smoke data.And a smoke dataset which covers a variety of scene including garages,gas stations and factories have been sorted out from the network.These videos are in the real fire video,including a variety of interference factors in the real scene.2.A smoke detection model based on improved CSP module(cross stage partial network)is proposed,and the comprehensive performance is the best in comparison with multiple models.Based on the complexity of smoke image texture features,this paper proposes to use multi-channel CSP module and multi-channel fusion CSP module to extract image features,in order to extract more different smoke texture features.Then these texture features are fused through the feature pyramid module to form a more representative abstract feature of smoke.This paper selects the optimal hyperparameters through experimental analysis,and through comparative experiments,it is proved that the performance of the model used in this paper in smoke detection exceeds the performance of the common general target detection model in this task.3.A method of accelerating smoke detection model reasoning by adding a smoke recognition sub network and a frame extraction network is proposed,which improves the model reasoning speed,reduces the computational power requirements of the model,and can separate and deploy the model to realize distributed reasoning.At present,the smoke detection model has high requirements for computing power,and can not carry out real-time smoke detection on hardware with poor computing power.In this paper,smoke recognition sub network and frame extraction network are used to end the reasoning process in advance,so as to speed up the model reasoning speed.At the same time,it reduces the hardware calculation requirements of the model.And the model is very suitable to be deployed separately on multiple hardware devices for distributed reasoning.This paper describes several application scenarios of the model in detail.4.A framework of dangerous event detection system is proposed,and a smoke detection prototype system is developed according to the framework.In this paper,the steps of different dangerous event detection are abstracted into different modules of the prototype system,and the framework of the prototype system is designed.Based on the framework and the smoke detection model trained in this paper,a smoke detection system is developed.The field test of the system verifies the effectiveness of the model and the robustness of the system,and shows the operation effect of the prototype system in practical application.