1 / 28

Device-free Localization based on CSI Fingerprints and Deep Neural Networks

Device-free Localization based on CSI Fingerprints and Deep Neural Networks. Rui Zhou Meng Hao, Xiang Lu, Mingjie Tang, Yang Fu University of Electronic Science and Technology of China (UESTC). Outline. Motivation Localization method Evaluations and comparisons Conclusions.

kim-spears
Download Presentation

Device-free Localization based on CSI Fingerprints and Deep Neural Networks

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript


  1. Device-free Localization based on CSI Fingerprints and Deep Neural Networks Rui Zhou Meng Hao, Xiang Lu, Mingjie Tang, Yang Fu University of Electronic Science and Technology of China (UESTC)

  2. Outline • Motivation • Localization method • Evaluations and comparisons • Conclusions

  3. Device-free localization • Indoor localization • Common solution: WiFi fingerprinting • Pinpoint mobile devices making use of RSSI • Device-free passive localization (DfL) • Objects do not carry mobile devices • Objects do not participate in localization process

  4. Device-free localization: how? Sensor-based Infrared, UWB, Ultrasound Specialized hardware High cost Vision-based Privacy issues Darkness Blind areas • WiFi sensing • Pervasive, low cost, device-free • No privacy issues, work anywhre • Through obstacles and walls • Under any light conditions, anytime • Basic rationale • Human movements alter signal propagation

  5. Channel State Information (CSI) • RSSI → CSI • Fine-grained information • Physical layer • Amplitude and phase of each subcarrier in a channel • OFDM: Orthogonal Frequency Division Multiplexing • CSI amplitude patterns are different at different locations → dependency • Amplitude of subcariers over time • Different colors represent different amplitude levels CSI at Location 1 CSI at Location 2

  6. Infrastructure of DfL • AP: Access point • Data transmission • Supporting 802.11n • Multiple TX antennas • MP: Monitoring point • Data retrieval • Supporting 802.11n • Multiple RX antennas • Server • Data processing • CSI sample • Amplitude of each subcarrier

  7. Previous methods • Signal propagation models • Difficult to establish accurate signal propagation models • Fingerprinting with classification • Bayesian algorithm • Support Vector Machines • Deep Neural Networks • Object location as the RP with most similar fingerprint • Fingerprinting with regression? • Localization is a continuous problem • Dependency between CSI fingerprints and locations

  8. The method • Method • Apply Deep Neural Networks (DNN) regression for localization, able to approximate arbitrary mathmatical functions • Apply Density-based Spatial Clustering of Applications with Noise (DBSCAN) for data denoising • Apply Principal Component Analysis (PCA) for contributing information extraction and dimension reduction • Comparisons • Investigate DNN structure: number of layers, number of neurons • Compare w. DNN classification, SVM classification, Bayesian • Compare w. state of the art Monostream, DeepFi

  9. Overview Training Localization

  10. CSI data collection • Raw training sample: <H, (x,y)> • His CSI matrix • Ntx ,Nrx: number of TX and RX antennas • Hij : CSI of TX i and RX j;Ns:number of subcarriers Hij=(h1, h2, ..., hk, ..., hNs) • hk: amplitude and phase of subcarrier k • CSI amplitude sample: Rij=(|h1|, |h2|, ..., |hk|, ..., |hNs|)

  11. CSI data denoising • Raw CSI data have noise • Apply DBSCAN to detect and reduce noise • Density-Based Spatial Clustering of Applications with Noise • Group points that are closely packed together • Mark as outliers that lie alone in low density regions

  12. CSI data denoising • Apply DBSCAN on each AP-MP pair • Regard each CSI sample as a point • Cluster CSI samples and detect noise • Distance between CSI samples:

  13. CSI contributing info extraction • A CSI sample has L = Ntx×Nrx×Ns dimensions • Apply PCA to extract contributing information, remove redundancy, reduce dimensionality • Principal Component Analysis • Find l (l < n) new features, each one being a linear combination of the original features, and • Clis cumulative contribution rate (CCR) of the first l features • Ccis predefined threshold of CCR (99%) • Dimensionality: 270  93

  14. Localization by DNN regression • Deep Neural Networks (DNN) regression • Features are CSI fingerprints after preprocessing • Target values are location coordinates • Model training is to establish a DNN, which represents the dependency between CSI fingerprints and location coordinates • Samples: CSI fingerprints labelled with coordinates of RPs • Goal: Structure of DNN, weights of neurons • Localization is to determine location coordinates according to CSI fingerprints through DNN • Samples: CSI fingerprints at unknown locations • Goal: location coordinates

  15. DNN structure Fully connected

  16. Localization by DNN regression • Activation function • Introduce nonlinearity into deep neural networks • Rectied Linear Units (ReLU) • Loss function • Measure the difference between ground truth and prediction • Mean distance error between true locations and estimated • Optimizer:Adaptive Moment Estimation (Adam)

  17. Evaluations - testbed 1 • Meeting room: 6 m x 6 m • 1 AP: 3TX, IWL5300 NIC • 1 MP: 3RX, IWL5300 NIC • Green circles: 40 RPs • Red stars: 12 arbitrary TPs • Sampling rate: 20Hz • Samples per location: 100 • Training samples: 4000 • Testing samples: 1200

  18. Evaluations - testbed 2 • Meeting room: 6 m x 6 m • 1 AP: 1TX • 1 MP: 3RX • Laboratory:6 m x 8 m • 1 AP: 2TX • 1 MP: 3RX • AP: TP-Link wireless router • MP: IWL5300 NIC • Green circles: 62 (40+22) RPs • Red stars: 32 (18+14) TPs • Training samples: 9300 • Testing samples: 4800

  19. Evaluations - training of DNN • Batch size: all • Epoch: 5000 • Learning rate: 0.001 • Meeting room:[64 128 256] • Training error starts at 4 m • 50 epoches: decrease slows down • 4800 epoches: converges on 0.058 m • Training time: 225 s (4000 samples) • Lab-meeting:[128 256 512 256 128] • Training error starts at 7 m • 120 epoches: decrease slows down • 3200 epoches: converges on 0.059 m • Training time: 1267 s (9300 samples)

  20. Evaluations - localization results • Regression time: • Meeting: 0.394 s • (1200 samples) • Lab-meeting: 1.257 s (4800 samples)

  21. Evaluations - preprocessing Effect of DBSCAN+PCA Data preprocessing with DBSCAN and PCA can improve accuracy.

  22. Evaluations - DNN layers & neurons • More complicated environments require more hidden layers • More numbers of layers or more numbers of neurons may not achieve better performance

  23. Evaluations - DNN layers & neurons Meeting: #Layers Meeting: #neurons Lab-Meeting: #Layers Lab-Meeting: #Neurons

  24. Evaluations - vs. classification Lab-meeting DNN classification SVM classification Bayesian algorithm Regression outperforms classification wrt. localization Meeting

  25. Evaluations - vs. other solutions Monostream: object recognition, device-free DeepFi: deep learning, device-based SVM regression: device-free Lab-meeting Deep learning outperforms traditional machine learning wrt. localization Meeting

  26. Evaluations - vs. classification and vs. other solutions

  27. Conclusions CSI is sensitive to environmental changes, thus capable of accurate device-free localization Deep learning demonstrates its effectiveness in localization, outperforming traditional machine learning Regression outperforms classification wrt. localization Due to small size of dateset, fully conntected DNN with a few hidden layers is adequate for accurate localization Data preprocessing with DBSCAN and PCA improves accuracy Evaluations achieved accuracy of single room 1.08 m and double rooms 1.50 m Open issue: mismatch of training and testing environments

  28. Thank you Questions?

More Related