Wifi-based Indoor Localization

학위논문 (박사)-- 서울대학교 대학원 : 기계항공공학부, 2016. 2. 김현진.With recent advances in smartphone industry, an indoor localization using smartphone becomes of increasing interests due to the need for indoor position information where GPS is not available. Fortunately, prevalence of wireless access points (APs), which are built in many buildings and public spaces, helps developing a wifi-based indoor localization without additional installation. This thesis considers a wifi-based indoor localization, where the wifi received signal strength (RSS) as a function of distance between a receiver (smartphone) and a transmitter (wireless AP) is applied to estimate both the floor level and position. The wifi RSS is non-linear and varying due to interference of the other radio signals and obstacles. Especially, signal attenuation and multipath effect are the major impediment against accurate localization. Because of those effects, the estimation methods using a propagation model of wifi RSS such as a triangulation and a least-square are inaccurate. This thesis proposes learning-based localization methods as a solution for those issues by training a nonlinear and unpredictable wifi RSS model. In particular, a semi-supervised learning algorithm is efficient for localization by removing a need for a large amount of the labeled training data. For example, in indoor localization, the labeled training data have to be collected manually. On the other hand, unlabeled data can be easily collected by recording wifi signal strength without the labels such as the position information and floor level. By using a large amount of unlabeled data and a small amount of labeled data, the semi-supervised learning algorithm improves efficiency and accuracy of the localization. The main contribution compared to the existing indoor localization can be found in i) mobile fingerprinting and ii) mapless localization. First, we address the efficiency for obtaining position training data, which is called fingerprinting. In the conventional fingerprinting, we have to collect the training data manually, which needs much time and effort of human. This thesis suggests a mobile fingerprinting based on a new semi-supervised learning algorithm, which provides the accurate pseudolabels of the unlabeled data. Second, by considering both privacy and communication issues between a service provider and user, this thesis proposes a mapless localization. With a concept of the crowdsourcing, we use the estimated locations obtained from the crowds, as the samples for learning a map. This is also based on a semi-supervised learning technique, and experimental results validate that more accurate map is learned as more participants join our localization system. Our final contribution involves field experiments in an office building at Seoul National University. We obtain training datapoints from different smartphone users who are not given any guideline about restricted attitude to carry the smartphone, for example, not to swing the smartphone or not to put it in pocket. From the experimental results, we find out that successful estimation of floor level and position. Also, the learned map is very close to the true map so that mapless localization is almost accurate as the result using the true map information.Chapter 1 Introduction 1 1.1 Motivation 1 1.2 Literature review 3 1.2.1 Wifi-based indoor localization 3 1.2.2 Semi-supervised learning 4 1.2.3 Mapless localization 6 1.3 Objectives and contribution 6 1.3.1 Pattern recognition of wifi RSS for indoor localization 7 1.3.2 Pseudolabelling for efficient training data collection 8 1.3.3 Mapless localization by trajectory learning 9 1.4 Thesis organization 11 Chapter 2 Feature Extraction from Wifi RSS data for Floor Classification and Landmark Detection 12 2.1 Characteristics of smartphone sensors 13 2.1.1 Inertial measurement unit (IMU) in smartphone 13 2.1.2 Wifi RSS characteristics 17 2.2 Semi-supervised learning for feature extraction 24 2.2.1 Generalized eigenvalue problem 24 2.2.2 PCA 25 2.2.3 FDA 26 2.2.4 Semi-supervised combination of FDA and PCA 27 2.3 Experimental results 29 2.3.1 Classification 29 2.3.2 Landmark detection 32 2.3.3 Effect of balancing parameter 34 Chapter 3 Mobile Fingerprinting and Pseudolabelling for Positioning 36 3.1 Learning based indoor localization 37 3.2 Basic semi-supervised optimization 38 3.2.1 Optimization framework 38 3.2.2 Laplacian embedded regularized least square (LapERLS) 43 3.3 Hodric-prescott filter optimization 46 3.4 Proposed semi-supervised optimization for pseudolabelling of mobile fingerprinting 47 3.5 Expereimental results of pseudolabelling 51 3.5.1 Sinusoidal trajectory 53 3.5.2 Overlapped trajectory 56 3.5.3 Revisiting the learned trajectory 56 3.5.4 Comparison with other semi-supervised learning algorithms 57 Chapter 4 Mapless Indoor Localization 59 4.1 Particle filter based localization framework 60 4.2 Gaussian process regression for modelling wifi RSS likelihood 63 4.2.1 Problem formulation 63 4.2.2 Hyperparameter selection 66 4.2.3 Building a prior 66 4.2.4 Experimental localization result 68 4.3 Map learning from a crowd 70 4.3.1 Concept of learning indoor trajectory 70 4.3.2 Trajectory learning algorithm 71 4.3.3 Experimental results of mapless localization 75 Chapter 5 Integrated Expereimental Result 77 5.1 Experimental setup 77 5.2 Result of map learning 79 5.3 Results of localization and floor classification 83 Chapter 6 Conclusion 90 References 92 Abstract (in Korean) 99Docto