Gait recognition using static activity-specific parameters

合集下载

Computer-Vision计算机视觉英文ppt

At the same time, AI MIT laboratory has attracted many famous scholars from the world to participate in the research of machine vision,which included the theory of machine vision , algorithm and design of system .
Its mainstream research is divided into three stages:
Stage 1: Research on the visual basic method ,which take the model world as the main object;
Stage 2: Research on visual model ,which is based on the computational theory;
the other is to rebuild the three dimensional object according to the two-dimensional projection images .
History of computer vision
1950s: in this period , statistical pattern recognition is most applied in computer vision , it mainly focuse on the analysis and identification of two-dimensional image,such as: optical character recognition, the surface of the workpiece, the analysis and interpretation of the aerial image.

智能机器人材料1

Abstract— Walking is one of the most common activities that we perform every day. Even if the main goal of walking is to move from one place to another place, walking can also convey emotional clues in social context. Those clues can be used to improve interactions or any messages we want to express. However, there are not many studies on the effects of the intensity of the emotions on the walking. In this paper, the authors propose to assess the differences between the expression of emotion regarding the expressed intensity (low, middle, high and exaggerated). We observed two professional actors perform emotive walking, with different intensities and we analyzed the recorded data. For each emotion, we analyzed characteristic features which can be used in the future to model gait patterns and to recognize emotions from the gait parameters. Additionally, we found characteristics which can be used to create new emotion expression for our biped robot Kobian, improving the human-robot interaction.

基于运动感知的帕金森患者冻结步态检测方法研究

华中科技大学硕士学位论文摘要冻结步态（Freezing of gait, FOG）是中晚期帕金森病（Parkinson’s disease, PD）患者最常见的步态障碍，极易导致患者跌倒并对其身心健康及生活质量产生严重的影响。

带有节律性听觉刺激的可穿戴系统可作为干预FOG的辅助工具，减少发病患者的跌倒风险。

冻结步态检测作为该系统的基础，还可为病情评估提供相关症状信息，具有重要的研究意义和应用价值。

本文设计了相关实验流程，采集了12名PD患者的冻结步态信号，其中运动传感器被部署在患者腰部、左右大腿等9个位置。

实验总共记录了2小时31分钟的数据，10名患者在试验期间表现出FOG，专业医师从视频记录中共识别出276个FOG事件。

在此基础上，本文围绕冻结步态检测过程中的划分时间窗、特征提取与选择、分类等问题展开了系统的研究。

本文对传感器各轴信号提取了13个典型时频域特征。

针对高维特征空间引起的分类精度下降和计算开销大等问题，选取互信息和方差分析这两种特征选择方法对特征重要性进行了评价并比较了两种方法的有效性。

本文评估了单一传感器和多传感器组合的检测效果，并使用方差分析为几种传感器配置选择了最佳特征。

在综合考虑检测性能、成本及实际部署要求等因素后，选择左小腿加速度计和陀螺仪的组合作为最佳传感器配置并为其选取了35个最佳特征。

本文将随机森林、AdaBoost、线性判别分析、多层感知神经网络应用于冻结步态分类算法设计中，研究了正负样本比例和时间窗大小对分类器性能的影响，逐步优化了检测模型的性能。

最终结果表明，当以1.25s时间窗、0.15s步长进行滑窗采样，提取左小腿加速度计和陀螺仪各轴的35个特征作为特征向量并使用由正负样本比例为1的样本集训练得到的AdaBoost模型进行分类时，可获得87.3%灵敏度，91.2%特异性，89.5%AUC的最佳检测效果。

关键词：帕金森病，冻结步态，传感器配置，特征选择，模式分类华中科技大学硕士学位论文AbstractFreezing of gait (FOG) is a common gait disorder among patients with advanced Parkinson's disease, which is associated with falls and negatively impact the patient's quality of life. Wearable systems with rhythmic auditory stimulation can be applied to help patients resume walking and reduce the risk of falls. As the basis of the system, the detection of freezing of gait can also provide relevant information for disease assessment, which has important research significance and application value.The thesis designed the relevant experimental procedures to obtain FOG signals from PD patients. Accelerometers and gyroscopes were placed on the patient's waist, left calve, right calve and the remaining 6 body parts. A total of 2 hours and 31 minutes of data was recorded in the experiment. Signals recorded from 10 patients with PD who presented the symptom of FOG and 2 patients who suffered from PD but they do not present FOG events. Professional physicians identified 276 FOG events from video recordings. On this basis, the research is carried out around segmentation, feature extraction, and classification.In this thesis, 13 typical time-frequency domain features were extracted from the sensor signals.Due to the low classification accuracy and high computational cost caused by high-dimensional feature space, mutual information and analysis of variance were selected to evaluate the importance of features and the effectiveness of the two methods were compared. The thesis evaluated the detection effects using several different configurations of sensors in order to conclude to the set of sensors which can produce optimal FOG episode detection and selected the best features for the optimal sensor configuration.After that, random forest, AdaBoost, linear discriminant analysis, and multi-layer perceptual neural network algorithm were applied to classification. The effects of the ratio of positive to negative samples and window size on classifier performance were studied, and the performance of the detection model was gradually optimized. The final results indicated that the proposed model was able to detect FOG华中科技大学硕士学位论文events with 87.3% sensitivity, 91.2% specificity, 89.5% AUC when using 1.25s time window (125 sample points) and 0.15s (15 sample points) step, the 35 features obtained from the gyro and accelerometer placed on patients’ left shank and AdaBoost classifier.Keywords: Parkinson’s disease, Freezing of gait, Sensor configuration, Feature selection, Pattern classification华中科技大学硕士学位论文目录摘要 (I)ABSTRACT (II)目录 (IV)第一章绪论 (1)1.1课题研究背景及意义 (1)1.2国内外研究现状及分析 (4)1.3本文工作及组织框架 (7)第二章冻结步态信号采集及样本集获取 (10)2.1基于运动感知的冻结步态检测过程及其检测性能指标 (10)2.2冻结步态信号采集 (12)2.3数据预处理 (16)2.4样本集获取 (18)2.5本章小结 (20)第三章冻结步态特征选择及传感器配置评估 (21)3.1冻结步态特征提取 (21)3.2特征选择方法的研究 (25)3.3传感器配置评估及特征选择 (31)3.4本章小结 (35)第四章冻结步态识别方法的研究 (36)华中科技大学硕士学位论文4.1分类算法性能的影响因素 (36)4.2分类算法设计 (39)4.3实验结果与分析 (46)4.4冻结步态识别结果与分析 (50)4.5本章小结 (51)第五章总结与展望 (52)5.1总结 (52)5.2展望 (54)致谢 (55)参考文献 (56)附录攻读硕士学位期间发表的论文与专利 (61)华中科技大学硕士学位论文第一章绪论1.1 课题研究背景及意义帕金森病（Parkinson's disease，PD）是由于中脑黑质多巴胺神经元大量变性死亡所引起的一种老年疾病[1]。

基于加速度传感器的人体运动状态识别研究.pdf

硕士学位论文基于加速度传感器的人体运动状态识别研究THE RESEARCH OF HUMAN ACTIVITY STATE RECOGNITION BASE ONACCELEROMETERS彭际群哈尔滨工业大学2014年12月国内图书分类号：TP399 学校代码：10213 国际图书分类号：621.3 密级：公开工学硕士学位论文基于加速度传感器的人体运动状态识别研究硕士研究生：彭际群导师：张春慨副教授申请学位：工学硕士学科：计算机科学与技术所在单位：深圳研究生院答辩日期：2014年12月授予学位单位：哈尔滨工业大学Classified Index: TP399U.D.C: 621.3Dissertation for the Master Degree in Engineering THE RESEARCH OF HUMAN ACTIVITY STATE RECOGNITION BASE ONACCELEROMETERSCandidate：Jiqun PengSupervisor：Associate Prof. Chunkai Zhang Academic Degree Applied for：Master of Engineering Speciality：Computer Science & Technology Affiliation：Shenzhen Graduate SchoolDate of Defence：September，2014Degree-Conferring-Institution：Harbin Institute of Technology摘要摘要随着社会的发展和互联网社交与娱乐的兴起，人们逐渐养成了上班不离电脑下班不离手机的生活方式，由于缺少科学的运动评估手段和有效的提醒方式，人们总是沉浸在工作和互联网社交与娱乐中，导致人们的运动量严重不足。

基于加速度传感器的人体运动状态识别，通过在身体上携带的传感器数据来识别人的运动状态，为人们运动的量化和自我健康认识提供了可能，本文描述了基于Android智能手机内置加速度感器的运动状态的识别，主要工作包括：提出了在不稳定加速度传感器设备上进行可靠数据采集的方法，通过设置异步缓存队列对数据进行重采样，保证数据频率保持在20HZ左右，以适应Android手机内置加速度传感器的不稳定性和不同机型之间的差异性。

基于强化学习的类人机器人步行参数训练算法

练算法。对步行参数进行降阶处理，利用强化学习算法优化参数，并设置奖惩机制。ＲｂｃｐＤ仿真平台上进行实验，在ｏｏｕ３结果证明了该算
法的有效性。关健词：类人机器人；步行参数；强化学习；奖惩机制
ＷａｋｎｇＰａａｅｅｓＴｒｉｎｇＡｌｏｉｈｍｆＨｕｍａｉｂｏｌｉｒｍｔｒａｎｉｇｒｔ０ｎｏｄＲｏｔＢａｅｎＲｅｎｆｒｅｅａｎｎｇｓｄ０ｉ．ｃｍｎｔＬｅｒｉ０
第３８卷第８期
Ｖ０－８ｌ３
・
计
算
机
工
程
２１０２年４月
Ａｐｒｌ０１ｉ２２
ＮＯ．８
ＣｏｐｅｍｕｔｒＥｎｇｎｅｉｇｉｅｒｎ
博士论文・
文章编号：ｌ０４８０２８彻ｌ—０文献标识码：ｏ —３２（１） — ３３ｏ２ｏＡ
２ＫｅａｆａｕｅｎｎｎｌｆｏｌｘＳｓｍｓｆｎｉｅｒｇＭｉｉｒｆｄｃｔｎＳｕｈａｔｉｅｓｙＮａｊｎ１０６Ｃｉａ．ｙＬｂｏＭｅｓｒｍｅｔｄＣｏ￣ｏｏＣｍｐｅｙｔＥｇｎｅｉ，ｎｓｙｏＥｕａｉ，ｏｔｅｓＵｎｖｒｉ，ｎｉｇ２０９，ｈｎ）ａｅｏｎｔｏｔ
ｒｉｆｒｅｅｔｌａｎｎｓａｐｌｄｔｐｉｉｅｔｓａａｔｒ，ｎｄｔｅｒｗａｄａｄｐｎｓｍｅｔｅｈｎｉｍｓｇｖｎｘｅｉｎａｅｕｌｈｅｎｏｃｍｎｅｒｉｇｉｐｉｏｔｚｅｏｍｈｅｅｐｒｍｅｅｓａｈｅｒｕｉｈｎｃａｓｉｉｅ．ＥｐｒｍｅｔｌｓｔｓｏｗｈｔｎｍｒｓｔａｔｅａｇｒｈｉｅｓｂｅｉｈｂＣｕＤｉｌｔｎｐｌｔｏｍ．ｈｌｏｉｍｓｆａｉｌｎｔｅＲｏｏｐ３ｓｍｕａｉａｆｒｔｏ

1、下载文档前请自行甄别文档内容的完整性，平台不提供额外的编辑、内容补充、找答案等附加服务。
2、"仅部分预览"的文档,不可在线预览部分如存在完整性等问题,可反馈申请退款(可完整预览的文档不适用该条件!)。
3、如文档侵犯您的权益，请联系客服反馈,我们会尽快为您处理(人工客服工作时间：9:00-18:30)。

GaitRecognitionUsingStatic,Activity-SpeciﬁcParametersAaronF.BobickAmosY.JohnsonGVUCenter/CollegeofComputingElectricalandComputerEngineeringGeorgiaTechGeorgiaTechAtlanta,GA30332Atlanta,GA30332afb@cc.gatech.eduamos@cc.gatech.edu

AbstractAgait-recognitiontechniquethatrecoversstaticbodyandstrideparametersofsubjectsastheywalkispresented.Thisapproachisanexampleofanactivity-speciﬁcbiometric:amethodofextractingidentifyingpropertiesofanindividualorofanindividual’sbehaviorthatisapplicableonlywhenapersonisperformingthatspeciﬁcaction.Toevaluateourparameters,wederiveanexpectedconfusionmetric—re-latedtomutualinformation—asopposedtoreportingapercentcorrectwithalimiteddatabase.Thismetricpre-dictshowwellagivenfeaturevectorwillﬁlteridentityinalargepopulation.Wetesttheutilityofavarietyofbodyandstrideparametersrecoveredindifferentviewingcondi-tionsonadatabaseconsistingof15to20subjectswalkingatbothanangledandfrontal-parallelviewwithrespecttothecamera,bothindoorsandout.Wealsoanalyzemotion-capturedataofthesubjectstodiscoverwhetherconfusionintheparametersisinherentlyaphysicaloravisualmea-surementerrorproperty.

1.IntroductionGaitrecognitionisasubﬁeldofbiometrics[8]andhastheadvantage(overotherbiometrics)ofbeingunobtrusivebe-causebody-invasivesensingisnotneededtocapturegaitinformation.Fromasurveillanceperspective,gaitrecogni-tionisanattractivemodalitybecauseitmaybeperformedatadistance,surreptitiously.Mostothermodalitiesrequireproximalsensing,makingitdifﬁculttoapplyunobservedandtomanypeople.Furthermore,humansexhibitthecapa-bilityofrecognizingpeoplefromimpoverisheddisplaysofgait[10,3,14]indicatingthepresenceofidentityinforma-tion.Inthispaperwedevelopagaitrecognition(orveriﬁca-tion)methodbaseduponstaticbodyandstrideparametersmeasuredduringwalking.Weﬁrstdiscusssomepreviousworkanddescribeseveralgeneraldeﬁcienciesinthoseef-forts.Wethendetailourapproachtoaddressthosecon-cerns.

1.1.PreviousworkPreviousworkinautomaticgaitrecognitionfromvisualmeasurementscanbedivided,roughly,intomodel-freeandmodel-basedapproaches.Bymodel-freewemeanthereisnounderlyingrepresentationofthethree-dimensionalstruc-tureofwalking;however,theydohaveanimplicitmodelofwalkingbuiltintotheirmethodsofextractingfeatures.Model-freeapproaches[7,11,12,1]analyzethemotionorshapesubjectsmakeastheywalk,andthefeaturesre-coveredfromthemotionorshapeareusedforrecognition.Model-basedtechniqueseithermodeltheperson[13,9]orexplicitlymodelthewalkofthepersonasitwillappearintheimagery[2].Inpersonmodels,abodymodelisﬁttothepersonineveryframeofthewalkingsequence,andparam-eters(i.e.angularvelocity,trajectory,orlimblengths)aremeasuredonthebodymodelasthemodeldeformsoverthewalkingsequence.Inwalkingmodels,amodelofhowthepersonmovesiscreated,andtheparametersofthemodelarelearnedforeveryperson.Mostoftheapproachestakentodatesufferfromthreekeydeﬁcienciesthatweseektoaddress.Oneisthelackofgeneralityofviewingcondition–notingtheexceptionsof[9,1].Second,mostresearchersdonotconsiderwhethertheconfusionsthatarisearecausedbyvisionyieldingnoisymeasurementsorbyhavingchosenfeatureswithlowdis-criminationpower.Perhapsthemostsigniﬁcantlimitationinmostpreviousworkingaitrecognitionisthemannerinwhichresultsarereported.Eventhoughthesizeofthedatabaseistypicallylessthantenpeople(sometimesasfewassix),resultsarereportedaspercentcorrect.Sucharesultgiveslittleinsightastohowthetechniquemightscalewhenthedatabasecon-tainshundredsorthousandsormorepeople.

1.2.OurapproachOurapproachtothestudyofgaitrecognitionattemptstoovercomethesedeﬁcienciesbytakingfourfundamentallydifferentstepsthanpreviousresearchers.1/M1/NNMx

P(x)Population density Pp(x)Individual uncertainty Pi(x)

Figure1:Uniformprobabilityillustrationofhowthedensityoftheoverallpopulationcomparestothetheindividualuncertaintyafterthemeasurementistaken.Inthiscasetheremainingconfu-sion—thepercentageofthepopulationthatcouldhavegivenrisetothemeasurement—is.

First,wedevelopagait-recognitionmethodthatrecoversstaticbodyandstrideparametersofsubjectsastheywalk.Ourtechniquedoesnotdirectlyanalyzethedynamicgaitpatterns,butusestheactionofwalkingtoextractrelativebodyparameters.Thismethodisanexampleofwhatwecallactivity-speciﬁcbiometrics.Second,asopposedtore-portingpercentcorrect,wewillestablishthereductioninuncertaintyofidentitythatoccurswhenaparticularmea-surementistaken.Third,todeterminewhetherresultingconfusionisbecauseofpoorvisionorpoorchoiceofadis-tinguishingfeature,wecomparevision-basedanalysistoasimilarmeasurementderivedfrommotion-capturedata.Becauseourstaticmeasurementsarebaseduponphysicalpropertiesofthepersonandthebehavior,wecancomputethosepropertiesdirectlyfromthree-dimensionallimbposi-tiondata.Lastly,wepresentanadhoccross-conditionmappingmethodthatallowsfortheidentiﬁcationofawalkingsub-jectviewedunderconditionsthataredifferentthanthoseatwhichtheirinitialdatawererecorded.Herewemeanadhocintheliteralsense:designedforthespeciﬁcsituation.