基于改进YOLOv8的隧道火灾检测研究

Research on tunnel fire detection based on improved YOLOv8

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个百分点,同时轻量化后模型参数减少了了29.7个百分点,向前推理时间降低了44个百分点；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。

Abstract:

The detection of tunnel fire presents challenges in terms of detection difficulty and the difficulty of deploying it directly on resource-limited embedded devices for real-time detection. In this paper, a tunnel fire detection algorithm based on an improved YOLOv8 is proposed. Firstly, a polarized attention mechanism is introduced to preserve high-resolution information and suppress redundant features while enhancing the capture of global information. Secondly, a novel Partial Convolution (PConv) is introduced to achieve low latency and high throughput in the model. Lastly, the WIoU (Weighted Intersection over Union) function is used to optimize the network"s bounding box loss, enabling fast convergence of the network. Experimental results demonstrate that the proposed network achieves a 1.3% improvement in mean Average Precision (mAP) on the utilized tunnel fire dataset. Furthermore, the lightweight model reduces the model parameters by 29.7%, and the forward inference time is reduced by 44%. The algorithm achieves a balance between accuracy and lightweight requirements, making it suitable for real-time detection in tunnel scenarios.