Real-Time Fire Detection Model Based on Attention Mechanism and Multi-scale Features Map

Jin Yang,Yi Liu,Weidong Fang,Gentuan Jia,Yunzhou Qiu

DOI: https://doi.org/10.1109/besc57393.2022.9995156

2022-01-01

Abstract:Fire detection and alarm trigger with the visual system are among the most significant tasks in constructing new intelligent cities. Nevertheless, existing fire and smoke detection algorithms are primarily limited to poor performance on precision and are not sensitive enough to detect small targets (i.e., flame and smoke). To tackle the problems, we propose a Fire-YOLOv5 model, which shows strong robustness and high accuracy for fire detection based on attention mechanism and multi-scale features map. The model introduces the optimized attention mechanism in the backbone network, which increases the weight of the network on the region of interest. Inspired by the idea of the bidirectional feature pyramid network, the partial path aggregation network is transformed into a bidirectional cross-scale connection to better integrate semantic features at each scale. In addition, we introduce small target detection layers to solve the problem of insensitive small targets in the video. Finally, the algorithm is designed with a majority voting mechanism for video frames, allowing for real-time warning during fire accidents’ initial stage. Fire-YOLOv5 can achieve a better balance between performance (93.5% Average Precision) and efficiency (21 Frame Per Second), especially meeting the requirements of accuracy and real-time detection for small targets.