Research On Real-time Detection Algorithm Of Lightweight Flame Image And Video Stream Based On YOLOv5

As an accident with high frequency and greater harm,fire seriously threatens the safety of people’s lives and property.Therefore,early detection technology for fires has become a key part of reducing the degree of damage.Early fire detection is mainly relied on hardware sensors,but sensor detection methods were greatly affected by factors such as environment and distance.The wide application of video surveillance equipment provided a new idea for flame detection,and image processing is applied to flame detection.In recent years,with the increasing development of artificial intelligence technology,flame detection methods based on deep learning have been more and more widely used.The current target detection algorithms are characterized by a wide variety and rapid development,but there are still a series of problems such as heavy weight,complex model,poor real-time performance and low accuracy.(1)In chapter III,an improved DSG＿YOLOv5 model is proposed to solve the problems of large volume,complex calculation,slow detection speed,and difficulty in deploying to mobile devices of flame detection models.Firstly,we call the k-means function to iterate over a priori box that fits our own dataset.Secondly,deep separable convolution is used to replace ordinary convolution in this model.Finally,the lightweight Ghost convolution is used to improve the C3 module,and the anchor calculation function is used to iterate out the prior box size suitable for this dataset.The results show that under the premise of ensuring accuracy,compared with the original YOLOv5 algorithm,the number of parameters and calculation of this model are reduced to 54.8% and 47%.The model size is reduced by 6.4M,and 5.47 more frames of data are processed per second,but the detection accuracy is decreased..(2)In order to solve the problem that the detection accuracy of DSG＿YOLOv5algorithm has decreased,a lightweight DGC＿YOLOv5 algorithm is proposed to improve the detection ability of the algorithm for small targets by introducing the Convolutional Block Attention Module In chapter IV.The m AP of the proposed DGC＿YOLOv5 algorithm on the test set can reach 94.4%,1.7 percentage points higher than that of the original algorithm,and the average detection speed on the video test set can reach 71 frames/s,which can meet the requirements of real-time detection.The amount of parameters and calculations are reduced to 41.2% and 34.8% respectively.and the model size is reduced by 8.4M,making it convenient for subsequent deployment on mobile devices.(3)In order to make the improved model practical,this paper uses the laboratory car as a platform and is equipped with the Kinect v1 camera to detect flame targets,and the detection accuracy of the DGC＿YOLOv5 model can reach about 86%.We directly call the HD video camera for experimental verification,and the average detection rate can reach 95.45 frames per second.Experiments show that the algorithm can achieve good flame detection accuracy and meet the requirements of lightweight and real-time detection.