Research Of Smoke Detection Via YOLOv5

Fire has brought huge economic losses to human society and caused irreparable damage to the natural environment on which we live.Therefore,a variety of fire detection technologies have emerged.At present,indoor fire detection technology is relatively mature,while outdoor fire has brought difficulties to fire detection due to the influence of various factors.With the development of artificial intelligence and computer vision technology,it is possible for visual camera equipped with artificial intelligence algorithm to detect fire.When a fire occurs,the first feature is smoke,so smoke detection research plays a very important role in fire warning and rescue.Among all kinds of fire detection methods,smoke detection technology based on visual camera has been widely concerned by various parties because of its wide applicability,flexible deployment,rapid detection and high real-time performance,and has a tendency to replace the traditional sensor-based device.Although considerable achievements have been made in smoke detection based on vision,there are still the following deficiencies in related research.First of all,the situations of fire occurrence are very diverse,which leads to different features such as the form,color and background of the smoke generated at the fire scene.However,the current public data set contains very limited scenes of smoke data,which leads to the insufficient ability of the smoke detection model to adapt to the real scene,and there is no unified and perfect data set to support the training model.Secondly,many smoke detection studies based on deep learning are carried out in stages.The suspected smoke area is extracted first and then detected,which fails to carry out real-time detection.The lightweight of the model is not considered when selecting the model.All kinds of problems restrict the application of smoke detection research to daily life.Based on the above problems,some targeted work has been done.The first step is to make a smoke data set,select an algorithm model with fewer parameters and faster detection speed to detect smoke,and improve the algorithm model to improve the detection accuracy of the model.The improved method does not increase the number of parameters of the model too much.The second step is to conduct experiments and analyze the experimental results.In the experimental design,the self-built smoke data set is compared with the open smoke data set,and the same algorithm model is trained.The detection effect shows that the self-built smoke data set is better than the open smoke data set.The improved model can also detect light smoke.The main work and innovation of this paper are as follows:(1)A smoke data set is made.In this paper,a smoke data set is created on the basis of the open smoke data set.The self-built smoke data set absorbs the advantages of each open smoke data set,including a variety of scene types and various interference factors,so that the trained model has a better detection effect on smoke.(2)Two improvement methods are proposed.One method is to use SIo U Loss as the frame loss function of the model,and the other method is to use SPPCSPC module in the model.The detection accuracy of the two improvement methods has been improved to a certain extent based on the YOLOv5 s algorithm model.The detection accuracy of YOLOv5 s algorithm model was increased by 1.28%.