基于人体骨架点的有生目标检测和行为预估  

Detection and behavior prediction of living aims based on human skeleton points

作　　者：张俊斌 景春阳 王希阔 蒋弘毅 王永娟[2] ZHANG Junbin;JING Chunyang;WANG Xikuo;JIANG Hongyi;WANG Yongjuan(Unit 63856 of PLA,Baicheng 137001,China;School of Mechanical Engineering,Nanjing University of Science and Technology,Nanjing 210094,China)

机构地区：[1]中国人民解放军63856部队,吉林白城137001 [2]南京理工大学机械工程学院,南京210094

出　　处：《兵器装备工程学报》2025年第1期221-229,共9页Journal of Ordnance Equipment Engineering

摘　　要：针对无人化作战平台在复杂战场难于正确判断有生目标以及目标威胁程度的问题,以HRNet为基线模型,构建人体骨架点检测模型评价指标,引入中继监督学习,并提出了改进通道剪枝方法,以实现有生目标的快速准确检测;采用自顶向下的骨架点组合方法,获得多目标场景下每个目标的行为估计;采集了典型动作下的人体姿态数据,获得行为特征的先验知识,采用多层感知机和径向基核支持向量机,训练获得人员目标行为分类器,并进行了试验验证。结果表明:有生目标的检测正确率和推理速度均得到了有效提高,行为预估准确率≥95%。In order to solve the problem in accurately judging the presence and threat level of living aims on unmanned combat platforms in complex battlefields,based on HRNet and the evaluation of human skeleton point detection models,a relay supervision mechanism is added,and an improved channel pruning methods are proposed,to achieve fast and accurate detection of living aims.Then,pose estimates for each target in multi-objective scenes are obtained by a skeleton-point combination method of top-down approach.Human pose data under typical actions are collected,and trained behavior classifiers using multi-layer perceptron and radial basis kernel support vector machine,and are verified by experimentations.The results show that the detection accuracy and inference speed of living targets have been improved,and the accuracy of behavior estimation is more than 95%.

关 键 词：无人化作战平台 有生目标 人体骨架点 姿态估计 行为预估 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]