基于改进YOLOv8的隧道火灾检测研究  

Research on Tunnel Fire Detection Based on Improved YOLOv8

作　　者：闵浩 屈八一[1] 谢子豪 MIN Hao;QU Bayi;XIE Zihao(School of Information Engineering,Chang'an University,Xi'an 710064,China)

机构地区：[1]长安大学信息工程学院,西安710064

出　　处：《计算机测量与控制》2024年第5期38-45,共8页Computer Measurement &Control

摘　　要：隧道内火灾检测存在检测困难和难以直接部署到资源有限的嵌入式设备进行实时检测的问题,提出一种基于改进YOLOv8的隧道火灾检测算法;首先引入极化注意力保持高分辨率信息来抑制冗余特征,同时增强全局信息的捕捉;其次引入了一种新的局部卷积PConv来实现低延迟和高吞吐量的模型;最后使用WIoU函数优化网络的边界框损失,使网络能够快速收敛。实验结果表明,该网络在所使用隧道火灾数据集上的平均精度mAP提升了1.3%,同时轻量化后模型参数减少了29.7个百分点,向前推理时间降低了44%;算法能够平衡精度和轻量化的需求,可以满足隧道场景下的实时检测。There are the difficulties of detecting fires inside tunnels and directly deploying to the embedded devices with limited resources for real-time detection,a tunnel fire detection algorithm based on improved YOLOv8 is proposed.Firstly,the polarized attention mechanism is introduced to preserve high-resolution information and suppress redundant features,while enhancing the capture of global information.Secondly,the novel partial Convolution(PConv)is introduced to achieve the model with low latency and high throughput.Finally,the WIoU function is used to optimize the loss of network bounding box,enabling the fast convergence of the network.Experimental results demonstrate that on the utilized tunnel fire dataset,the mean average precision(mPA)of the proposed network improves by 1.3%.Furthermore,the model parameters of the lightweight model reduces by 29.7%,and the forward inference time by 44%.The algorithm meets the requirements of accuracy and lightweight,making it suitable for real-time detection in tunnel scenarios.

关 键 词：YOLOv8 局部卷积 WIoU 极化注意力 轻量化 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]