Lawrence W.C. Wong Ambient Intelligence Laboratory Interactive & Digital Media Institute National University of Singapore

Transcription

1 Indoor Localization Methods Lawrence W.C. Wong Ambient Intelligence Laboratory Interactive & Digital Media Institute National University of Singapore 1

2 Background Ambient Intelligence IDMI Established in April of 8 labs initially started in IDMI (Interactive and Digital Media Institute) Seed funding from NUS to kick-off Received NRF-IDM funding of about $5.8M for research into: Localization techniques Autonomous networks Sensor and wireless networks Context-aware platforms 3-D scene analysis and modelling Embodied media interactions (virtual world real world interactions) Virtual Mirror World Real World 2

3 Traditional Methods Triangulation Range and/or direction from/to reference points Fingerprinting 2 stages, i.e. offline (training) and online (run-time use) Dead-reckoning Device-centric and portable Hybrid Proximity sensing Combination of methods

4 Triangulation Based on geometric properties of triangles to determine location Require either range and/or direction information Range based: Received Signal Strength (RSS) Time of Arrival (ToA) Time Difference of Arrival (TDoA) Direction based: Angle of Arrival (AoA) Angle of Departure (AoD)

5 Triangulation Range based Received signal strength Signal path attenuation determines range from reference point Time of Arrival (ToA) Propagation delay determines range from reference point Time Difference of Arrival (TDoA) Overcomes the need for absolute time sync Correlation analysis provides a time delay t i t j corresponding to path difference to receivers i and j

6 Triangulation Direction based Angle of Arrival (AoA) / Angle of Departure (AoD) Use acoustic arrays or RF antenna arrays Estimate phase differences or relative signal amplitudes Require coherent receiver 2 1 AP 2 AP 1 Triangulation with AoA

7 Fingerprinting Methods RF-based scene analysis Offline stage collect features (fingerprints) Online stage estimate location by matching online measurements against previously collected fingerprints 5 common approaches: Probabilistic k Nearest Neighbour (knn) Neural networks, Support vector machine (SVM) Smallest M-vertex polygon (SMP)

8 Fingerprinting Probabilitic Assume there are n location candidates L 1, L 2, L 3,..., L n, and s is the observed signal strength vector during the online stage Decision rule is to: Choose L i if P(L i s) > P(L j s), for i, j = 1, 2, 3,..., n, and j i. Applies for discrete location candidates Can be extended to any estimated 2-D location ( xˆ, yˆ ) by interpolating the position coordinates ( xˆ, yˆ) n i 1 P( L i s)( x L i, y L i )

9 Fingerprinting k Nearest Neighbour Use online RSS to search for k closest matches of known locations in signal space Weighted or unweighted approaches Fingerprinting Neural Networks RSS & corresponding location coordinates used to train neural network After training, appropriate weights are obtained Multilayer perceptron (MLP) network with one hidden layer is used Input layer bias may be used

10 Fingerprinting - SVM A technique for data classification and regression SVM data fusion location algorithm can reduce estimation errors SVM constructs a hyperplane or set of hyperplanes in a high dimensional space, thereby facilitating classification, regression or other tasks Good separation is achieved by hyperplane that has the largest distance to the nearest training datapoints Maximum-margin hyperplane and margins for a SVM trained with samples from two classes. Samples on the margin are called support vectors

11 Fingerprinting - SMP Uses online RSS values to search for candidate locations in signal space w.r.t. to each signal transmitter separately The generalized weighted L p distance between a measured RSS vector [x 1 x 2 x N ] and a database entry [x 1 x 2 x N ] is given by: L p 1 N N i 1 1 w i x i x ' i p 1/ p Choose M closest database entries (those with the smallest signal distance) and estimate the location based on the average of the coordinates of these M points

12 Dead Reckoning Less dependent on installed infrastructure Start with known initial position Estimate displacement and direction Error accumulation over time and distance travelled Usually require accompanying alternative error control and reduction techniques

13 Proximity Sensing Range based sensing Does not detect location, only spatial closeness Uses proximity sensors that can be either: Electrostatic Electromagnetic Ultrasonic Common technologies: RFID (Radio Frequency IDentification) Bluetooth WiFi Cellular mobile Sonar

14 Problems & Challenges Non Line-of-sight (LOS) propagation Signal scattering Interference, particularly for EM based techniques Vulnerability to environment changes Labour intensive Fine resolution usually require high bandwidth Computational complexity Cost Scalability: Different environments Large areas

15 Channel Information Based Fingerprinting Dimension of RSS-based location fingerprint is too small New fingerprint is based on Amplitude of Channel Impulse Response (ACIR) with: logarithmic transformation to accentuate feature differentiation local smoothing to average out spurious randomness of closely separated locations Non-parametric Kernel Regression is used for location estimation

16 Fingerprinting Theoretical Analysis Probabilistic framework that covers: Probability of error distance conditioned on online feature vector Region of Confidence (RoC) Asymptotic error performance Non-parametric Density Estimation techniques employed RSS fingerprint used for comparative validation Probabilistic Model: N training samples, (c n, s n ), n = 1, 2, N, where s n = [s n,1, s n,2,, s n,m ] T are the training RSS vectors c n = [x n, y n ] T corresponding coordinates Online samples comprise: s = [s 1, s 2,, s M ] T is the online RSS vector c = [x, y] T is its corresponding coordinates From fingerprint system: c ~ x ~, y~ is estimated coordinates T Derive joint and conditional PDF of location error

17 Fingerprint Theoretical Analysis Error vector is: Let = magnitude of error vector = angle of error vector Then, we have: The joint PDF f c,s (c,s) can be estimated using Non-parametric Density Estimation techniques Then the conditional PDF f s ( s) can be determined: For being less than a distance r 0, the RoC is expressed as: T y y x x c c e ~, ~ ~ T T y y x x ] ~ ~, [ ] sin, cos [ ) ( r s d s f s r P P ) (, ] sin ~, cos ~ [ 2 0, s f d s y x f s f s T s c s

18 Fingerprint Theoretical Analysis Testbed Set-up 3 APs 125 training points 126 testing points KNN Probabilistic Emphirical Predicted Emphirical Predicted 25% RoC (m) 50% RoC (m) 75% RoC (m) Ave Error (m)

19 SparseTrack: Indoor Pedestrian Tracking with Sparse Infrastructure Emergence of personal hand-held devices with low-cost sensors (digital compass, accelerometers) Fingerprint approach incur high setup cost and vulnerable to environment changes Trilateration approach require at least 3 reference points and susceptible to signal scattering degradations Dead reckoning (DR) approach is straightforward but suffers from error accumulation Hybrid approach : Combine simplicity of DR with occasional corrections in sparse infrastructure

20 SparseTrack - Methodology DR with step meter using range sensors comprising: Accelerometers Digital compass Region of uncertainty grows in DR over space and time Path of mobile device occasionally gets close to reference Beacon Nodes (BNs) Proximity to BN reduces uncertainty region Generalized fusion algorithm via Gradient Descent iteration arrives at Maximum Likelihood solution

21 SparseTrack - Performance SparseTrack is able to track users much better than DR alone, with reductions in average error by up to 71.2%. SparseTrack is robust and works well even when initial device location is unavailable and range updates are intermittent. Path 1 Path 2 Path 3 Path 4 Ave Error of DR (m) Ave Error of SparseTrack with Initial Location (m) 0.27 (71.2%) 0.32 (56.8%) 0.41 (44.6%) 0.28 (59.6%) Ave Error of SparseTrack w/o Initial Location (m) 0.36 (62.3%) 0.37 (50.0%) 0.5 (31.5%) 0.38 (45.8%) Average Rate of Correction (per second)

22 Estimation of Center of Mass Displacement based on Gait Analysis Objective: Ambulatory algorithm to estimate Center of Mass (CoM) displacement Approach: Estimation of CoM displacement based on gait analysis and segmental kinematics Human lower body is modeled as a kinematic chain of rigid segments linked by joints, and pelvis is considered as root joint in the model Toe joint at zero displacement from ground is selected as reference for kinematic transmission Performance of algorithm is evaluated using simulated sensor measurements obtained from an optical motion capture database Squared norm [(m/s 2 ) 2 ] Squared norm [(m/s 2 ) 2 ] Left foot detected stance phase squared norm of acc detected stance phase sample index,t Right foot detected stance phase squared norm of acc detected stance phase sample index,t Z-axis (m) X-axis (m) 2 0 CoM level Position TRUE EST INT sample index,t 2 0 TRUE EST INT sample index,t

23 Cluster-Based Localization - Background Localization in ad hoc networks Organize into hierarchical ad hoc network with cluster heads as reference nodes Each node can detects distances from other nodes within its communication range Nodes are assumed to have their own location initially with some error Aim to derive and track location of all nodes in network through a distributed iterative approach

24 Cluster-based Localization - Methodology Cluster-based anchorless localization algorithm 2 phases: Primary Phase use uploaded information from members to derive initial positions of members within each cluster and the results are broadcasted to the cluster members Refinement Phase every node use least-square minimization to get more accurate result of itself Extended Kalman filter used in Primary Phase Performance compared against DV-distance algorithm with different node density and number of reference nodes

25 Multiple Sensor Fusion Multiple DR systems carried by a user have stable relative displacements w.r.t. center of motion, and therefore to each other Robust tracking based on a generalized maximum a posteriori (MAP) sensor fusion approach Step based DR: Orientation Projection for Arbitrary Device Posture Noise Filtering Step Detection Stride Length Estimation Heading Orientation 2 sets of sensors are carried with arbitrary but fixed orientations Reasonably stable relative displacement with respect to each other

26 Multiple Sensor Fusion Constrained optimization: Maximize f (ˆ l l a, l b a l a ) f (ˆ l b l b ) subject to l where l a and l b are the locations of the sensors, and R is the radius of uncertainty Optimal l a and l b on the line segment that connects lˆa and lˆb with distance R away from the middle point of that line segment a l b 2R

27 Web Services Architecture to support Context-Aware Applications Why Web Services? Different technologies (Web Services, CORBA, RMI) Interoperation among heterogeneous software systems Accessible on the Web at a given URL Distributed composition of services (Web services orchestration) Fixed Service Contract Reuse, Modularity Typical Services Real-time Location Tracking Systems Proximity-based Location Services (Bluetooth, Ekahau) Cricket-based Location Tracking Systems Typical Applications Real world Virtual world interaction Social tagging and networking Environmental sensing and monitoring

28 Service-based Software Infrastructure Supports standard Web services protocols, such as HTTP (REST Representational State Transfer) & SOAP. Context-aware applications may or may not be Web services. Current Aggregation of Web Services is done in a static approach Introduces the concept of service gateway called POEM Service. POEM Service provides following features: Integrates multiple service Service Orchestration 3 internal components: Mapping across different coordinate systems Fusion of multi-modal location sensing technologies Output generation of contextmetadata content

29 POEM Service Gateway Services implemented: Ekahau location service Bluetooth location & proximity service Cricket mote location service SOAP Web Service Communication Restful Web Service Communication Context-aware Application Bluetooth Proximity Service Poem Service (Service Gateway) Ekahau Proximity Service SOAP Web Service Communication

30 POEM Portal Web Interface Poem Portal Registration of App object Id (binding virtual object id with physical object id) in Poem Service via Poem Portal App Developer/Admin Poem DB Restful Web Service Communication Context-aware Application Poem Service (Service Gateway)

31 APIs for Context-Aware Applications API References are published in Poem Portal APIs are RESTful web services Request URLs can be easily constructed Allow binding objects through API. Applications need XML parser. Sample Web Service API (Get Nearby Users) This API allows the application to retrieve the nearby users by giving the user id. Request URL Request parameters appid Id servicetype method

32 On-going & Future Work Fusion techniques for multi-modal location sensing Localization in MIMO systems Vision-based ranging and localization POEM Service: Dynamic Service Orchestration Service Discovery Scalability Load Balancing Data fusion Data Visualization (Map) Common Coordinate Systems for different type of location Systems.

33 End of Presentation