【作者】 黄康；

【导师】 郑利平；

【作者基本信息】 合肥工业大学 ， 计算机技术（专业学位）， 2020， 硕士

【摘要】 不时发生的火灾给社会带来严重的损失,对火灾的预防、监控和灾害发生时对人群疏散的研究已经成为热点。由于空间和视角的限制,比较大规模火灾发生时,处于现场的救援人员往往无法准确得知周边情况。为了更好辅助救援人员抢险救灾和人群疏散工作,本文提出一种基于航拍图像的大范围路网提取算法以及火灾发生时疏散路网更新生成方法,利用深度学习网络模型和图像处理技术将航拍图像中的道路、火情信息生成可以理解的语义信息,从高空视角给出宏观总体态势信息,辅助应急管理、人群疏散等决策工作。本文主要研究内容如下:1)综述了灾害周边环境感知所用到的道路识别算法和火焰识别算法研究现状,并对现有的深度学习神经网络模型和算法进行了分析。2)提出一种道路识别网络模型D-Cross Link Net,设计了跨分辨率交叉连接、双空洞卷积模块,实现从航拍图中抽取城市道路网络任务,通过实验证明了结果的有效性和优越性。3)实现了综合RGB、YCb Cr、HSI三种颜色空间规则的火焰识别算法,识别并定位图像中火灾点,生成标记效果图,适用于多种格式和不同分辨率的图像。将道路识别结果和火焰识别结果融合,将道路上的火灾点标记为障碍点,生成具有火情标记的道路网络,服务于应急疏散工作。更多还原

【Abstract】 The occurrence of fire disaster brings serious losses to humans.The research on fire prevention,monitoring and crowd evacuation has become a hot topic.Due to the limitation of space and perspective,when a large-scale fire occurs,rescuers are often unable to know the real-time environmental conditions around the fire site,which made the rescue work more difficult.In order to help people to do rescue work and crowd evacuation task,this thesis proposes a large-scale road extraction method based on aerial images and the generation of evacuation road networks in fire disaster,using deep learning networks and image processing techniques to compare the roads and fire site information in aerial images.Our work provides understandable semantic information to rescuers from a high-altitude perspective,assist in decision-making work such as emergency management and crowd evacuation.The main research work of this thesis is as follow:1)The research status of road network extraction are reviewed,the existing fire detection algorithms are analyzed and summarized in this thesis.2)A road extraction network D-CrossLinkNet is proposed,it use cross-resolution connection and double dilated convolution blocks to achieve the task of road networks extraction from aerial images.Experimental results on two benchmark datasets(Beijing and Shanghai dataset,Deep Globe dataset)demonstrate the effectiveness and superiority of the proposed network.3)A fire recognition algorithm based on RGB,YCbCr and HSI space rules is implemented,which can be used for images of various formats and resolutions to identify and locate fire site in the images.The results of road network and fire recognition is combined.We mark the fire sites on the road as obstacles,thereby generating a road network with fire site marks.更多还原