基于双分支头部解耦和注意力机制的灾害环境人体检测  

Pedestrian detection method in disaster environment based on double branch Decoupled Head and Attention Mechanism

作　　者：郝帅[1] 杨晨禄 赵秋林 马旭[1] 孙曦子 王海莹 孙浩博 吴瑛琦 HAO Shuai;YANG Chenlu;ZHAO Qiulin;MA Xu;SUN Xizi;WANG Haiying;SUN Haobo;WU Yingqi(College of Electrical and Control Engineering,Xi’an University of Science and Technology,Xi’an 710054,China)

机构地区：[1]西安科技大学电气与控制工程学院,陕西西安710054

出　　处：《西安科技大学学报》2023年第4期797-806,共10页Journal of Xi’an University of Science and Technology

基　　金：国家自然科学基金项目(51804250);中国博士后科学基金项目(2020M683522);陕西省科技计划项目(2021JQ-572,2020JQ-757)。

摘　　要：灾害环境中,利用计算机视觉可以有效协助消防员进行救援,缩短搜救时间。针对受灾人体目标受多尺度、部分遮挡以及环境干扰导致传统算法难以准确检测的问题,提出一种基于双分支头部解耦和注意力模型的灾害环境人体检测网络。首先,为解决灾害环境下小尺度人体目标造成的漏检问题,在YOLOv5框架下,构造浅层检测层以增强网络对小目标识别能力;其次,针对灾害环境中人体目标易淹没在复杂背景中进而导致目标特征无法有效表达的问题,通过融合轻量化注意力模块以增强人体目标的显著度,并在特征的原始输入和输出节点间添加连接以提高网络多尺度特征融合能力;最后,为了减少人体检测网络中分类和回归任务的差异性对检测性能造成的影响,构建双分支头部解耦检测器分别用于人体目标的识别和定位。为验证所提算法的优势,在多种灾害救援场景下进行测试验证,并与5种经典算法进行比较。相较于对比算法,所提算法精度最高,平均精度和召回率分别可达92.2%和90.5%,不仅能够准确检测出人体目标,而且具有良好的实时性和鲁棒性。Computer vision can facilitate the resue of firefighters in a disaster with the searching time shortened.To solve the problem that the traditional algorithm is difficult to accurately detect the human body target in a disaster environment due to multi-scale,partial occlusion and environmental interference,a human body detection network based on decoupled head and attention model is proposed.Firstly,for the missing detection caused by small-scale human body targets in disaster environment,YOLOv5 framework was used to construct a shallow detection layer to enhance the recognition ability of the network for small targets.Secondly,aiming at the problem that human targets are prone to submerge in a complex background in a disaster environment,which leads to the inability to effectively express the target features,the lightweight attention module was fused to enhance the saliency of human targets,and the links were added between the original input and output nodes of features to improve the multi-scale feature fusion capability of the network.Finally,in order to reduce the influence of the differences between classification and regression tasks on the detection performance in the human detection network,a decoupled head was constructed for human target recognition and localization respectively.And the advantages of the proposed algorithm have been verified in various disaster rescue scenarios over those with five classical algorithms.Compared to the comparison algorithm,the proposed algorithm has the highest accuracy,and the mean avearage precision and recall rate can reach 92.2%and 90.5%respectively.It can not only accurately detect human targets,but also has good real-time and robustness.

关 键 词：深度学习 人体检测 多尺度检测 注意力机制 解耦检测器 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]