【作者】 李方；

【导师】 王铁成； 佟为明；

【作者基本信息】 哈尔滨工业大学 ， 电机与电器， 2012， 博士

【摘要】 随着全球定位系统（GPS）在室外定位中的成功应用，对定位服务在人们日常活动的室内场所得以广泛开展的渴求日渐突出。在现有室内定位技术中，位置指纹法由于定位成本低、方法实现灵活而逐渐成为研究的热点。然而，室内无线信号的时变性也给位置指纹法实现精确定位带来很大的难度。因此，提高位置指纹法在室内复杂环境下的定位性能和实用性不仅具有重要的理论价值，而且也成为加快对其由理论研究向实际应用转变的关键。本文以ZigBee无线网络为应用平台，对位置指纹法完成定位所需的两个阶段中涉及的相关技术进行研究。在分析国内外研究现状及进展的基础上，指出了位置指纹法目前存在的三大问题，即构建位置指纹数据库的工作量问题、位置指纹的匹配效率问题和室内定位与跟踪的实现方法问题。然后，从分析信号强度作为场景特征所表现出的数据特性入手，针对以上问题分别提出相应的解决方法。为分析信号强度作为场景特征所表现出的数据特性，利用ZigBee网络环境中测得的信号强度样本，采用理论分析与实验验证相结合的手段对数据特性的表现形式和成因进行了研究。在此基础上构建了一个用于估计位置指纹法平均定位误差的概率模型，并分析了采样位置间距、网络接入设备数量和位置等因素对位置指纹法定位性能的影响，从而为定位系统实际部署过程中相关参数的合理选取提供一定的理论指导。为降低离线阶段信号强度采样的工作量，提出了一种基于空间变异理论的位置指纹数据库构建方法。给出了一个典型的信号强度样本变异函数计算流程。提出了一种基于加权最小二乘法的信号强度理论变异函数拟合算法。根据拟合得到的模型趋势项和随机项反映出的信号强度在空间上的连续性和变异性关系，采用普通克里金技术对待估位置信号强度进行最优线性无偏估计。实验结果表明，该方法对信号强度的估计精度好于距离反比加权算法，基于该方法建立的位置指纹数据库可在降低采样工作量的同时保证定位精度。为提高位置指纹的匹配效率，提出了两种基于改进k-means算法的位置指纹聚类方法。分析了传统k-means算法对位置指纹分类效果不佳的原因，并从寻找新的位置指纹二次特征提取方法及采用“软划分”技术两方面加以改进。其中，FC-ID-FKM法将位置指纹归为区间值数据，在由区间中值和大小张成的特征空间中，利用模糊k-means算法对其进行聚类。FC-KFKM法将位置指纹归为一种服从正态分布的区间值数据，通过区间中值和大小确定的正态分布函数将位置指纹映射为特征空间中的一点,并在该空间中采用基于核方法的模糊k-means算法对其进行聚类。实验结果表明，基于两种方法得到的位置指纹集合聚类趋势更为明显，对位置指纹的分类效果好于k-means算法。为解决室内复杂环境下的定位实现问题，提出了一种基于D-S证据理论的位置指纹室内定位方法。构造了信度分配的mass函数，建立了一个基于pignistic概率的信度分配冲突程度评价指标。根据指标大小，采用信度打折的证据预处理方法消除定位过程中出现的合成悖论现象。为降低信任区间重叠时的决策风险，根据合成后焦元获得的信任区间确定分布函数，将两焦元信度在并集区间上降序排列的概率作为最终的决策依据。实验结果表明，该方法具有收敛速度快，决策风险低的优点。在证据间高冲突时的定位精度好于贝叶斯推理法。此外，针对基于位置指纹法的室内跟踪问题，提出了一种改进的粒子滤波跟踪算法。建立了二维平面内跟踪问题的状态空间模型，并根据本文提出的信号强度估计方法完成粒子权值的求解。实验结果表明，该方法对实际移动路线的跟踪精度好于KNN法和Kalman滤波器。更多还原

【Abstract】 With the successful applications of global positioning system (GPS) onoutdoor positioning, the demands of indoor positioning services are getting shaper.Because of the low positioning cost and flexible realization, the locationfingerprint method (LFM) is becoming the research focus in the existing indoorpositioning techniques. However, the time variation of indoor wireless signalmakes it difficult for LFM to realize pinpoint. So, improving the positioningperformance and practicability of LFM in the complex indoor environment notonly has important theoretical value, but also becomes the key of accelerating thetransition from theoretical study to application. In this paper, related techniques ofLFM in the two positioning stages are studied in the ZigBee network. Based onanalyzing research status and development of LFM at home and abroad, it ispointed out that the problem of construction workload of location fingerprintdatabase, the problem of matching efficiency of location fingerprints and theproblem of indoor positioning and tracking realizations are the three mainproblems of LFM. And then, starting with analyzing signal strength characteristics,methods for resolving these problems are proposed respectively.In order to analyze the data characteristics of signal strength as the scenefeature, the manifestations and causes of characteristics are studied by theoreticalanalysis and experimental verification with actual ZigBee signal strength samples.And then, a probability model is given to estimate the average positioning error ofLFM. Based on the model, the relationships between positioning performance ofLFM and space of sampling locations and numbers and locations of network accessdevices are studied, which can provide theoretical guide for configuring the relatedparameters reasonably during the positioning system deployment.In order to decrease the sampling workload of signal strength in the off-linestage, a method for constructing location fingerprint database is proposed based onthe spatial variability theory. A typical construction procedure of signal strengthsample variogram is given according to the actual signal strength samples. A fittingmethod of signal strength theory variogram is proposed based on the weightedleast square method. According to the continuity and variability of signal strengthat different locations reflected by trend and stochastic component of the mode, thebest linear unbiased estimation of signal strength is achieved by ordinary kriging.The experimental results show that the proposed method can get better estimationprecision of signal strength than weighted distance inverse method. The location fingerprint database constructed by it can decrease the sampling workloadefficiently and ensure the positioning precision at the same time.In order to improve the matching efficiency of location fingerprint, twoclustering methods for it are proposed based on the improved k-means algorithm.The methods raise the clustering accuracy by using the new further featureextraction methods of location fingerprint and elastic classification technologiesafter analyzing the problems of k-means algorithm application on locationfingerprint classification. FC-ID-FKM regards location fingerprint as a kind ofinterval-valued data and adopts fuzzy k-means algorithm to cluster it in the featurespace established by interval median and size. FC-KFKM summarizes locationfingerprint as a kind of interval-valued data obeyed normal distribution and adoptsfuzzy kernel k-means algorithm to cluster it in the feature space established bynormal distribution function determined by interval median and size. Theexperimental results show that the location fingerprint sets obtained by theproposed methods have remarkable clustering trends. They can all get betterresults on location fingerprint clustering than k-means.In order to realize the accurate positioning in the complex indoor environment,a Dempster-Shafer based indoor positioning method is proposed. The massfunction is constructed to allocate the belief and a pignistic probability basedevaluation index is given which can reflect the conflict degree on belief allocation.According to the index value, the evidence pretreatment method by belief discountis adopted to eliminate the fusion absurdity during positioning. To decrease thedecision risk on overlaped belief intervals, the distribution functions of focalelements are constructed according to the belief intervals and the probabilitiesranking in a descending order on union of belief intervals are calculatedrespectively which are taken as the basis of decision-making finally. Theexperimental results show that the proposed method has fast convergence and lowdecision risk. The positioning accuracy of the method is better than it of Bayesianinference on the high conflict degree. In order to realize the LFM based indoortracking, an improved particle filter tracking algorithm is put forward. The statespace model for tracking is established in the two-dimensional plane and theparticle weight is calculated according to signal strength estimation methodproposed in the paper. The experimental results show that the proposed method hasbetter tracking effect than KNN and Kalman filter.更多还原