【作者】 李飞；

【导师】 李德信；

【作者基本信息】 西安理工大学 ， 机械制造及其自动化， 2018， 硕士

【摘要】 随着智能机器人技术的快速发展及其在制造系统和人机交互领域的应用,智能化的人体动作识别技术为满足智能系统高柔性需求提供了有效途径。为了对操作人员动作的有效识别,需提取能够完整准确描述动作的人体动作特征,实现人体动作的识别分类,本文通过对人体不同动作特征、动作识别模型、参数优化方法的深入分析,对多特征进行融合,建立了动作识别模型,并对模型参数进行优化,实现了对人体动作的有效识别。基于人体关节的三维空间坐标,通过特征融合描述人体姿态和动作,建立了行为动作模型。在人体骨架模型简化的基础上,建立了行为动作模型,提取计算、融合了人体主要关节角度、速度以及相对位置三种互补特征,通过融合特征描述行为姿态,用姿态系列表示动作。结合人体动作视频数据库,为便于计算,以最多特征维数样本为标准,对数据库中每个动作样本特征进行傅里叶插值,使得动作样本的特征维数一致,对样本数据进行归一化预处理。最后,利用主成分分析法提取特征主要成分,从而降低特征维数,减少冗余信息。基于融合特征,构建了四种多分类动作识别模型,包括一对一多分类模型、一对多多分类模型、有向无环图多分类模型以及决策树多分类模型。依据多分类模型中不同的核函数进行行为识别测试,选择测试识别精度最高的核函数作为多分类模型的核函数,实现对测试动作样本的识别。对比分析不同模型的识别结果,得到在多分类识别模型中,一对多多分类模型取得较高识别率。在分析多分类识别模型参数的基础上,分析得到影响识别准确率的重要参数。利用改进的网格优化算法、萤火虫优化算法以及狼群优化算法,对该多分类模型的参数进行优化,提高了多分类识别模型的识别率,通过实验仿真对比得到在狼群优化时,识别速度较快、准确率高,且鲁棒性强。更多还原

【Abstract】 With the rapid development of intelligent robot techology and its application in the field of manufacturing systems and human-computer interaction,intelligent human motion recognition technology provides an effective way to meet the high flexibility requirements of intelligent systems.In order to realize the effective recognition of the operator’s actions,it is necessary to extract human action features that can completely and accurately describe the actions and realize the recognition and classification of the human actions.This article provides an in-depth analysis of different human movement characteristics,motion models,and parameter optimization methods.in this paper,the multi-features were fused and a motion recognition model was established.The parameters of the model were optimized to realize effective recognition of human motion.Based on the three-dimensional coordinates of human joints,the human body poses and movements are described through feature fusion,and a behavioral action model is established.Based on the simplification of the human skeleton model,a behavioral action model was established,and three complementary features of the human main joint angle,speed,and relative position were extracted and combined.The behavioral gesture was described through the fusion feature and the gesture was used to represent the action.In combination with the human motion video database,for ease of calculation,the Fourier transform is performed on each motion sample feature in the database with the most characteristic dimension sample as the standard,so that the feature dimensions of the motion samples are the same,and the sample data is normalized.Finally,principal component analysis is used to extract the main components of the feature,which reduces the feature dimesion and reduces redundant information.Based on the fusion feature,four multi-classification action recognition models were constructed,including one-to-one multi-classification model,one-to-many multi-classification model,directed acyclic multi-classification model,and decision tree multi-classification model.According to the different kernel functions in the multi-classification model,the behavior identification test was performed.The kernel function with the highest test and recognition accuracy was selected as the kernel function of the multi-classification model to realize the identification of test action samples.By comparing and analyzing the recognition results of different models,it is found that in the multi-classification recognition model,the one-to-many multi-classification model achieves a higher recognition rate.On the basis of analyzing the parameters of multi-class identification model,important parameters affecting the recognition accuracy rate are obtained.Using the improved grid optimization algorithm,firefly optimization algorithm,and wolves group optimization algorithm,the parameters of the multi-classification model are optimized,and the recognition rate of the multi-classification recognition model is improved.The speed of recognition in the optimization of wolves is optimized through experimental simulation and comparison.Faster,accurate,and robust.更多还原