Aiming at the problems of missing detection and misdetection,large number of model parameters and difficult location in fire detection,a lightweight fire detection and ran-ging method based on binocular vision and improved YOLOv8n was proposed.Pictures were taken by binocular camera,and the improved detection algorithm YOLOv8n-AEM and the existing ranging algorithm SGBM were used for detection and ranging.Firstly,variable ker-nel convolution AKConv and EMA attention mechanisms are introduced into the backbone network to effectively extract fire features by constructing irregular convolutional nuclei.Then,the C2f-SCConv module is constructed in the neck network to reduce the model param-eters and improve the detection speed through feature recombination.Secondly,the loss func-tion is improved based on the minimum point distance to solve the problem of missing detec-tion and false detection caused by overlapping fire source and light source.Finally,the detec-tion head of small target is added to improve the detection ability of small flame.The experi-mental results show that the improved detection algorithms P,R and mAP are 83.6％,76.4％and 83.6％respectively,which are improved by 2.5％,3.6％and 4.8％respectively.The parameter number and model size were 2.54 M and 5.1 MB,which decreased by 15.3％and 15％,respectively.The accuracy error of ranging is less than 2.5％,which proves that the improved method can accurately complete the fire detection and ranging.
关键词
火灾检测/双目视觉/测距/YOLOv8n/轻量化
Key words
fire detection/binocular vision/distance measurement/YOLOv8n/light weight
维普
万方数据