面向消防需求的火焰快速检测方法

Firefighting-oriented Rapid Flame Detection Method

消防系统中的火焰检测设备在执行检测任务时存在精度低、实时性差等问题,会导致误报率较高,发现不及时,从而错过施救的最佳时机。为了解决上述问题,以YOLOv7-tiny算法为基础,对ELAN-tiny结构进行改进,设计ELAN-TP结构,该结构大量减少了原结构的冗余计算。为提高主干网络的特征提取能力,在主干网络的末端引入ConvNeXt结构。使用SimAM注意力机制对颈部网络结构进行改进,增加模型对目标的定位精度以及对不同尺寸特征的融合能力。结果表明:改进模型的火焰检测精度mAP@50达到了57.7%,相较于原模型提升了4.8%,FPS(检测帧率)达到了97,提升了13。将改进后的模型在边缘计算设备Jet⁃son Nano上部署,FPS在20上下波动,满足了实时性的需求,能够较好地在各种场景下执行火焰检测任务,具有较高的实用价值。

The edge computing devices in the fire protection system have problems such as low accuracy and poor real-time performance when performing flame detection tasks,which can lead to high false alarm rates and missed opportunities for timely rescue.To address the above issues,based on the YOLOv7-tiny algorithm,the ELAN-Tiny structure is improved and the ELAN-TP structure is designed to significantly reduce the redundant computation of the ELAN-tiny structure.To improve the feature extraction capability of the backbone network,the ConvNeXt structure is introduced at the end of the backbone network.The neck model is improved by using the SimAM attention mechanism to increase the model’s positioning accuracy for detecting targets and its ability to integrate features of different sizes.The results show that the improved model achieves a flame detection accuracy of 57.7% at mAP@50,an increase of 4.8% compared to the original model,and a detection frame rate of 97,an increase of 13.After deploying the improved model on the edge device Jetson Nano,the FPS fluctuates around 21,which meets the real-time requirements and enables better execution of flame detection tasks in various scenarios,with high practical value.