应急救援是针对各类突发事件而采取预防、预备、响应和恢复的活动与计划，各类突发事件常伴随火灾的发生，火灾破坏力极强、影响范围甚广、严重威胁人类的生命财产安全，火灾给人类带来的损失远高于其他突发事件。为了保障应急救援时的应急通信，课题组目前研发了一套宽窄带融合的应急通信指挥系统，该系统中的应急布控球尚缺乏烟火检测功能。早期的传感器型烟火检测器成本较大、误报率高，近年来基于深度学习的烟火检测算法极大地改善了烟火检测的性能，但是存在着检测精度低、鲁棒性差等问题，因此设计实现一种基于应急布控球的改进烟火检测算法有着重要的意义。

本文首先针对烟火数据集图像质量差，对比度低等问题，采用高斯滤波降噪和直方图均衡对烟火图像进行预处理，提升了烟火数据集的图像质量。同时为了减少疑似烟火的静态目标对烟火检测的影响，采取五帧帧差法进行动态目标提取，并提出改进Canny边缘检测与上述算法进行“与”融合，获取边缘更加清晰完整的动态目标轮廓，提高了烟火检测算法的检测精度。

其次针对YOLOv5（You Only Look Once，YOLO）烟火检测模型泛化能力差、检测准确率低、重叠目标检测效果较差等问题，研究实现了基于YOLOv5烟火检测模型的多阶段改进。首先在输入端采用Mosaic-9数据增强和标签平滑处理，提高了模型的泛化能力并防止过拟合的发生；其次在特征融合层加入融合CBAM（Convolutional Block Attention Module，CBAM）和ECA-Net（Efficient Channel Attention Networks，ECA-Net）注意力模块，突出图像中的烟火目标，同时采用Kmeans++聚类算法提高先验框与目标框的匹配度，提高了烟火检测的准确率；最后采取软化非极大值抑制（Soft Non Maximum Suppression，Soft-NMS）改善了重叠烟火目标的检测效果。经实验分析，改进后的烟火检测模型的平均精度值高达93.69 %，较原始YOLOv5提升了5.6%。

最后采取Deepstream框架实现改进烟火检测模型在Jetson Nano嵌入式主板的部署，同时实现了基于DeepStream底层框架Gstreamer的语音告警，最后与应急布控球进行衔接，对系统整体功能进行测试分析，烟火检测准确率高于93%，误报率低于2%，语音告警性能稳定无误差，极大地提高了烟火检测准确率，减少应急救援中烟火灾害造成的各项损失。

Emergency rescue is the prevention, preparation, response and recovery activities and plans for all kinds of emergencies. All kinds of emergencies often associated with the occurrence of fire, fire is extremely destructive, affects a broad scope, a serious threat to human life and property security, the harm of fire to human beings is much higher than other emergencies. In order to guarantee the emergency communication during the emergency rescue, an emergency communication command system with wide and narrow band fusion has been developed at present, which lacks the fireworks detection function in the emergency deployment ball-control. The early sensor-based fireworks detector are costly and has false alarm rates. In recent years, the fireworks detection algorithm based on deep learning has greatly improved the performance of fireworks detection, but there are some problems such as low detection accuracy and poor robustness. Therefore, it is of great significance to implement an improved fireworks detection algorithm based on emergency deployment ball-control.

In this paper, aiming at the problems of poor image quality and low contrast of the fireworks dataset, which uses Gaussian filtering noise reduction and Histogram equalization to preprocess fireworks images to improve the image quality of fireworks dataset. At the same time, in order to reduces the impact of static objects suspected to be fireworks on fireworks detection, the Five-frame frame difference method is adopted for dynamic target extraction, and the improved Canny edge detection algorithm is proposed to "and" fuse with the above algorithm.The algorithm can obtain clearer and more complete dynamic target contour to improve the accuracy of fireworks detection.

Secondly, in view of the problems of poor generalization ability, low detection accuracy, and poor detection effect of overlapping targets of the YOLOv5 fireworks detection model, the multi-stage improvement based on the YOLOv5 fireworks detection model was realized. Firstly, Mosaic-9 data enhancement and Label smoothing were carried out at the input to improve the generalization ability of the model and prevent overfitting. And the fusion CBAM and ECA-Net attention module are added to the feature fusion layer to highlight the fireworks targets in the image, at the same time, Kmeans++ clustering algorithm is used to improved the matching degree between prior boxes and target boxes, and the accuracy of fireworks detection is improved. Finally, Soft-NMS is adopted to improve the detection effect of overlapping fireworks targets. Through experimental analysis,compared with the original YOLOv5, the average precision of the improved fireworks detection model is increased by 5.6% to 93.69%.

Finally, the DeepStream framework is adopted to realize the deployment of the improved fireworks detection model in Jetson Nano embedded motherboard, and the voice alarm based on DeepStream underlying framework Gstreamer is implemented.Connected with the emergency deployment ball-control,the overall function of the system is tested and analyzed, fireworks detection accuracy is greater than 93%, false positives are less than 2%, and voice alarm accuracy is 99%,which greatly improves the accuracy of fireworks detection and reduces various losses caused by fireworks disasters in emergency rescue.