Characteristics Based Fire Detection System Under the Effect of Electric Fields With Improved Yolo-v4 and ViBe

To address slow image-based detection of fires under the effect of electric fields and its limitation to static fire characteristics, this paper proffersa video-based fire detection system with improved you only look once version 4 (Yolo-v4) and visual background extractor (ViBe) algorithms. The proposed system uses a simplified weighted bi-directional feature pyramid network (Bi-FPN) in place of the path aggregation network (PANet) as a feature fusion network in Yolo-v4. Using multiple dynamic fire characteristics, it can eliminate falsely detected frames. The ViBe algorithm is improved to consider the sudden change of light triggered by fire flickering. Compared with other fire detection algorithms, the proposed system achieves 98.9% fire detection accuracy with a false detection rate of 2.2%. It can extract target fires by adjusting to sudden changes of light using no more than 16 frames. Moreover, the system achieves fire detection with more dynamic fire characteristics compared with the image-based fire detection under the effect of electric fields.