基于ResNet50对地震救援中人体姿态估计的研究
信息技术与网络安全 3期
邬春学,贺欣欣
(上海理工大学 光电信息与计算机工程学院,上海200093)
摘要: 调查发现,地震中死亡人数增加的原因主要是错过救援的黄金时间,因此可通过救援无人机自动对受灾人员进行行为识别与状态分析。人体姿态估计是指对图像中人体关节点和肢体进行检测的过程,在人机交互和行为识别应用中起着重要的作用,然而由于背景复杂、肢体被遮挡等因素导致标注人体关节点和肢体十分困难。因此提出一种结合ResNet50及CPM的模型,该模型通过获取图像特征和精调机制,计算出关节点依赖关系,最后划分到对应人体。实验表明,该模型与其他模型对比能够提高复杂场景下人体姿态估计的效果。
中圖分類號: TP391
文獻標(biāo)識碼: A
DOI: 10.19358/j.issn.2096-5133.2022.03.009
引用格式: 鄔春學(xué),賀欣欣. 基于ResNet50對地震救援中人體姿態(tài)估計的研究[J].信息技術(shù)與網(wǎng)絡(luò)安全,2022,41(3):50-58,70.
文獻標(biāo)識碼: A
DOI: 10.19358/j.issn.2096-5133.2022.03.009
引用格式: 鄔春學(xué),賀欣欣. 基于ResNet50對地震救援中人體姿態(tài)估計的研究[J].信息技術(shù)與網(wǎng)絡(luò)安全,2022,41(3):50-58,70.
Research on human posture estimation in earthquake rescue based on ResNet50
Wu Chunxue,He Xinxin
(School of Optical-Electrical and Computer Engineering,University of Shanghai for Science and Technology, Shanghai 200093,China)
Abstract: It was found that, the main reason for such a high number of deaths lies in the missing of prime rescue time. So rescue UAV can be used to recognize the behaviors of affected population automatically and analyze their status. Human pose estimation refers to the process of detecting humans′ joints and limbs in image, which plays a crucial role in human machine interaction and application of action recognition. However, due to the factors such as complex background and covering of limbs, it is very difficult to note the human joints and limbs in image. To address the issue, this paper proposed a model combining ResNet50 and convolutional pose machine(CPM). According to the model, image features are obtained by residual network and the dependence between joints is obtained by fine adjustment mechanism. Finally the key points aggregated are divided to the corresponding human body. Experiment shows that compared with other human pose estimation models, such model can enhance the effect of human post estimation under complex earthquake rescue scenario.
Key words : neural network;human pose estimation;ResNet50;part affinity fields;earthquake rescue
0 引言
據(jù)EM-DAT報道[1]稱,2000年至2019年間特大地震自然災(zāi)害導(dǎo)致死亡的受災(zāi)人數(shù)在九種自然災(zāi)害死亡人數(shù)中居首位,大約占總受災(zāi)人數(shù)的58%,在地震發(fā)生后高效率地救援十分必要。基于成熟的硬件設(shè)備[2],救援無人機搜尋傷員對其進行動作識別與狀態(tài)分析,可顯著提高救援的效率。因此,開展基于深度學(xué)習(xí)的實時無人機災(zāi)后救援人體姿態(tài)估計研究顯得十分必要[3-4]。
目前,無人駕駛的多旋翼無人機配備了高清攝像頭和高性能的電池,可滿足長時間懸停并傳輸震后實時救援的畫面[5]。在此基礎(chǔ)上,通過無人機獲取的震后救援現(xiàn)場的實時圖像,采用深度學(xué)習(xí)檢測和跟蹤方法[6]對受災(zāi)后傷員的位置以及人體姿態(tài)進行檢測,以供指揮中心進行快速部署救援并能夠掌握震后的全局狀況。通常情況下,其對人體骨骼的關(guān)鍵部件的具體檢測精度有一定的要求,不僅要對整個人體進行精準(zhǔn)檢測,而且還要對人體的關(guān)鍵節(jié)點,例如頭部、肩關(guān)節(jié)、肘關(guān)節(jié)、膝蓋等部分進行更詳細(xì)的檢測和跟蹤,從而掌握更詳細(xì)的震后人員狀態(tài)。
本文詳細(xì)內(nèi)容請下載:http://m.ihrv.cn/resource/share/2000004026
作者信息:
鄔春學(xué),賀欣欣
(上海理工大學(xué) 光電信息與計算機工程學(xué)院,上海200093)

此內(nèi)容為AET網(wǎng)站原創(chuàng),未經(jīng)授權(quán)禁止轉(zhuǎn)載。
