Ultra-Wideband Geo-Regioning A Novel Pattern Recognition Localization Technique

Christoph Steiner and Armin Wittneben Communication Technology Laboratory, ETH Zurich

Motivation I • Clustering and/or coarse localization of UWB nodes – Logical grouping of nodes in close vicinity • Spatially correlated sensor data => distributed source coding

– Cluster position gives coarse location information of sensors • Location aware applications and tracking

• Low complexity UWB nodes – Complexity / energy consumption constraints for sensor nodes – Shift complexity to receiver (access point, cluster head)

2

Motivation II • Indoor and rich multipath environment – Warehouse, Office Building, Industry Hall

• Robust to non-LOS situations – ToA localization approaches break down

• Pattern recognition localization method based on UWB channel impulse response (CIR) fingerprinting – “UWB Geo-Regioning”

3

UWB Channel: Enabling Multipath Resolution

4

Localization Methodology \ Scenario •

Close node (TX) positions relative to TX Ù RX distance – Propagation environment does not change significantly – Typical temporal multipath pattern

•

RX

Multipath Reflections Region B

A-priori knowledge at RX – Pattern definition – Learn patterns for all regions

•

TX

Region C Region A

Tracking of mobile TX possible

5

Probabilistic CIR Modeling • Region characterized by probability model for CIRs • Gaussian tap assumption • Band-limited and sampled CIRs in equivalent baseband representation are modeled as proper jointly Gaussian random vector • CIR (

) model from region A:

6

Probability Models for Regions • A-priori knowledge specifies region model parameters • Maximum likelihood parameter estimates: – N i.i.d. observations per region

7

Findings and Conclusions • Accounting for correlated CIR taps increases reliability and robustness of UWB Geo-Regioning – Correlation is a result of alignment procedure and propagation environment – Correlation is region specific

• Low error probability with reasonable receiver complexity – – – – –

1 GHz BW 5 ns observation window CIR estimation SNR of 25 dB 10 x 10 covariance matrix On average 99% of UWB nodes are assigned correctly

• Channel estimation at RX must provide full CSI at high sampling rate • Amount of required a-priori knowledge determines effort for site survey 8

Future Work • Channel Impulse Response Models – Alternatives to Gaussian taps

• Influence of 9 Bandwidth 9 Temporal observation window size – Amount of a-priori knowledge • Blind Geo-Regioning • Asilomar submission

• Application in sensor networks – Correlation model for spatially correlated sensors readings – Application to distributed source coding 9

10

Pattern Recognition Algorithms • Binary hypothesis testing problem – Classification of CIR ( ) with unknown region

• Maximum likelihood decision rule Hermitian matrix => analytical error probabilities in paper

• Independent tap assumption (reduces complexity) – Average power delay profile (APDP) represents region

11

Agenda • Introduction – Motivation and Localization Methodology

• Characterization of Regions – Definition of Patterns

• Pattern Recognition Algorithms • Performance Evaluation – Based on Measured Data

• Summary and Conclusions

12

CIR Measurement Campaign • Rich multipath environment – Cellar room

• Frequency range 3 – 6 GHz – Limited by hardware

• • • • •

620 CIRs per region 22 regions Minimum region distance ¼ 10 cm Maximum region distance ¼ 16 m 2 representative region pairs – LOS and non-LOS: 13 and 18 – LOS and LOS: 11 and 13

• Region dimensions: 27 cm x 56 cm 13

Performance Evaluation • 620 measured CIRs divided into 434 a-priori CIRs and 186 CIRs for classification/testing – 70% Ù 30% – Required amount of a-priori knowledge is not investigated here

• Cross Validation approach – Random separations of a-priori and test CIRs – Averaged results for Pe|A and Pe|B (binary problem)

14

Signal-to-noise Ratio (SNR) • Performance evaluation for different SNR operating points and two algorithms – Independent taps (APDP) – Full second order statistics (COVARIANCE)

• Zero mean, additive white Gaussian noise • Updated probability model: • SNR definition

15

Performance Results

•

COVARIANCE approach shows significant SNR gains and improved robustness – Correlations among CIR taps are present and help to distinguish regions 16