PaperlayerView exact revision on arXiv

Source-linked AI summary

CSI-based Fingerprinting for Indoor Localization: A Deep Learning Approach

Xuyu Wang, Lingjun Gao, Shiwen Mao, Santosh Pandey

arXiv:1603.07080v1cs.NI

TL;DR

Indoor localization requires accurate positioning despite complex indoor propagation and the limitations of RSS fingerprints. DeepFi uses CSI-derived deep-learning fingerprints trained offline with a greedy layer-wise procedure, then estimates locations online with RBF-based probabilistic fusion. Experiments in two indoor environments show lower location errors than existing RSSI- and CSI-based schemes.

  • Problem

    Indoor localization must provide accurate, low-complexity, fast positioning despite multipath, shadowing, fading, and delay distortion.

  • Method

    DeepFi trains a four-hidden-layer network on CSI and uses its weights as fingerprints, with greedy layer-wise training and RBF-based probabilistic online localization.

  • Results

    DeepFi outperformed several existing RSSI- and CSI-based schemes in living-room and computer-laboratory experiments.

  • Takeaways & Limitations

    CSI-based deep-learning fingerprints can reduce localization error across the two evaluated indoor environments.

Abstract

from arXiv · show

With the fast growing demand of location-based services in indoor environments, indoor positioning based on fingerprinting has attracted a lot of interest due to its high accuracy. In this paper, we present a novel deep learning based indoor fingerprinting system using Channel State Information (CSI), which is termed DeepFi. Based on three hypotheses on CSI, the DeepFi system architecture includes an off-line training phase and an on-line localization phase. In the off-line training phase, deep learning is utilized to train all the weights of a deep network as fingerprints. Moreover, a greedy learning algorithm is used to train the weights layer-by-layer to reduce complexity. In the on-line localization phase, we use a probabilistic method based on the radial basis function to obtain the estimated location. Experimental results are presented to confirm that DeepFi can effectively reduce location error compared with three existing methods in two representative indoor environments.

I. INTRODUCTION

Indoor localization must handle complex radio propagation while maintaining high accuracy, low complexity, and short online processing time. DeepFi addresses these requirements with CSI-based deep-learning fingerprints and a two-phase localization system.

  • Indoor localization faces multipath, shadowing, fading, and delay distortion, while mobile devices require accurate, low-complexity, fast positioning.
  • Fingerprinting uses offline training to build a database and online testing to estimate an unknown position.
  • RSS-based localization is simple but suffers from temporal variability and coarse information that does not exploit individual subcarriers.
  • DeepFi uses CSI to capture amplitude and phase information for each subcarrier and antenna, providing finer-grained wireless-channel data than RSS.
  • DeepFi trains CSI fingerprints with a four-hidden-layer deep network and uses the trained system for online localization.
  • The system is evaluated in living-room and computer-laboratory environments against existing RSSI- and CSI-based schemes.

B. Hypotheses

DeepFi’s CSI hypotheses concern temporal stability, location-dependent subcarrier clusters, and antenna diversity. These properties motivate using all normalized CSI amplitudes in deep-learning fingerprints rather than averaging them.

  • CSI amplitude values are hypothesized to remain more stable over continuously received packets at a fixed location than RSS values.
  • 90% of CSI standard deviations were below 10% of the average, compared with 60% for RSS, across 150 sampled locations.
  • CSI amplitude values are hypothesized to form different numbers of subcarrier clusters across locations because of multipath and channel fading.
  • Measurements from 50 living-room locations identified dominant clusters of subcarriers with similar CSI amplitudes, supporting cluster-based location features.
  • The three Intel WiFi Link 5300 antennas are hypothesized to provide different CSI features that increase training and test-sample diversity.
  • DeepFi collects 90 CSI measurements per packet from three antennas and 30 subcarriers, using all normalized amplitudes instead of averaging antennas.
  • Feature-based fingerprints store deep-learning weights for locations, capturing CSI characteristics and reducing noise during localization.

B. Weight Training with Deep Learning

DeepFi trains a four-hidden-layer deep network and stores its learned interlayer weights as location fingerprints. Greedy layer-wise learning, followed by reconstruction-based fine-tuning, reduces training difficulty and yields weights for online localization.

  • The deep network has four hidden layers with decreasing widths, satisfying K1 > K2 > K3 > K4.Pretraining, unrolling, and fine-tuning are used in sequence.
  • DeepFi represents each training location by the weights between connected network layers, using normalized CSI magnitudes as input.The fingerprint includes W1, W2, W3, and W4 for the four successive layer connections.
  • A greedy algorithm trains the deep network layer by layer with a stack of RBMs because maximum-likelihood learning is difficult for the complex model.Earlier-layer parameters are frozen while samples from conditional probabilities train subsequent layers.
  • CD-1 updates each RBM’s weights using hidden-state and reconstructed-state products, with α as the step size.The update uses hi−1hi − ˆhi−1ˆhi for weights and corresponding differences for biases.
  • Fine-tuning minimizes reconstruction error between input v and reconstructed data ˆv, producing optimal weights stored as fingerprints.The network is unrolled with forward propagation before back-propagation adjusts the layer weights.

C. Location Estimation based on Data Fusion

DeepFi estimates locations by matching online CSI data against fingerprints learned offline, then applying a Bayesian model with an RBF likelihood. Multiple packets are fused, and the final estimate is a weighted average of reference locations.

  • Bayesian location estimation: After offline training, DeepFi tests unseen positions using a probabilistic model based on Bayes’ law.The paper motivates probabilistic methods as performing better than deterministic ones.
  • Offline training: The training algorithm learns eight weight matrices for each of N training locations and stores them as fingerprint groups.Training uses m packet receptions with 90 CSI values for each training location.
  • RBF likelihood: The likelihood Pr(v|Li) is modeled as a Gaussian radial basis function derived from the input and reconstructed CSI data.The formulation uses the input data v, reconstructed input data v̂, variance σ, and coefficient of variation λ.
  • Data fusion: Multiple packets are grouped into batches and their RBF values are averaged to improve location estimation and accelerate matching.The online input contains n packet receptions, each with 90 CSI values, alongside fingerprint groups for known training locations.
  • Location output: The mobile device’s position is estimated as a weighted average of all reference locations using the resulting probabilities.The Bayesian model assumes a uniform prior Pr(Li) = 1/N across the N reference locations.

A. Experiment Methodology

The experiments use a TP-Link router and a Dell laptop with an IWL 5300 NIC to collect CSI, then evaluate DeepFi’s training and online matching procedures. The study examines how training-spot spacing, access-point count, and indoor environment affect localization.

  • Testbed: The testbed combines a TP-Link router access point with a Dell laptop equipped with an IWL 5300 NIC for CSI collection.The laptop stores raw CSI values in the NIC firmware, accessed through a modified Linux setup.
  • Experiment procedure: During training, CSI collected at each location is converted into learned features and stored as fingerprints for online matching.Test data are matched to the closest training spot with a similar feature.
  • Training-spot spacing: Sparse training spots can reduce localization accuracy because a measurement may have similarly strong similarity to many random spots.The paper links this ambiguity to fuzziness during the test phase.
  • Access-point configuration: DeepFi can achieve precision similar to Horus and FIFS with one access point, although its offline training requires more time and computation.The method searches CSI features for the most matched training spot.
  • Indoor environments: The evaluation compares location errors across environments, including an open room with unobstructed center paths and a more complex environment with fewer LOS paths.The open-room setting is a 4 × 7 m2 room with 50 uniformly scattered positions and one access point placed at one end.

1) Living Room in a House:

The paper presents layouts for training and test positions in the living room and laboratory environments, with network structure specified for the living-room setting and laboratory scenario.

  • The living-room layout is used to designate training and test positions.
  • The laboratory layout is likewise used to designate training and test positions.
  • The living-room DeepFi network uses hidden-layer sizes K1 = 300, K2 = 150, K3 = 100, and K4 = 50.
  • The laboratory scenario uses 50 training positions and 30 test positions, with a network structure of K1 = 500, K2 = 300, K3 = 150, and K4 = 50.

2) Computer Laboratory:

The comparison evaluates DeepFi against three existing localization methods using the same measured dataset for all schemes.

  • DeepFi is compared with FIFS, Horus, and Maximum Likelihood using the same measured dataset.

3) Benchmarks and Performance Metric:

The paper defines mean sum error as the comparison metric and presents localization-error results for the living-room experiment.

  • Mean sum error E measures localization performance from estimated and actual user positions across K locations.
  • Table I reports mean errors for the living-room and laboratory environments.
  • Fig. 9 shows the CDF of localization errors in the living-room experiment.

B. Localization Performance

DeepFi is evaluated across indoor environments and parameter settings, showing lower localization errors with CSI-rich inputs while exposing accuracy–processing-time trade-offs.

  • Localization Performance: 0.95 meters and 1.8 meters are DeepFi’s mean distance errors in the living room and laboratory, respectively.DeepFi outperforms FIFS, whose corresponding mean errors are 1.2 meters and 2.3 meters.
  • Localization Performance: About 60% of living-room test points have DeepFi errors under 1 meter, while over 60% of laboratory test points have errors at 1.7 meters.
  • Localization Performance: In the laboratory, tables obstruct line-of-sight paths and magnify multipath, weakening the correlation between signal strength and propagation distance.DeepFi exploits CSI subcarriers and performs well in this non-line-of-sight environment.
  • Impact of Different Antennas: More than 60% of test points in the 90-CSI scheme have errors under 1.5 meters, compared with fewer than 40% for the 30-CSI single-antenna scheme.Mean error decreases from around 2.12 meters to about 1.84 meters when using three antennas.
  • Impact of Different Antennas: 2.3 seconds versus around 2.5 seconds is the average processing time for single-antenna and three-antenna schemes over 100 packets.The three-antenna scheme takes about 10 percent extra processing time and achieves a 15 percent improvement in localization precision.
  • Impact of the Number of Test Packets: 1.83 meters versus 1.93 meters is the mean localization error for 300 versus 5 test packets.Increasing the number of packets tends to reduce mean error and standard deviation.
  • Impact of the Number of Test Packets: 4.2 seconds versus about 1.7 seconds is the execution time for 300 versus 5 test packets.The paper prefers fewer packets because additional packets slightly improve precision while increasing collection and processing time.
  • Packet Processing: Packets are divided into equal-sized batches so they can be processed in parallel during the test phase.

3) Impact of the Number of Packets per Batch:

DeepFi’s batch-size study finds that increasing packets per batch mainly reduces execution time, while averaging across batches supports robust localization. The broader CSI experiments examine environmental stability and training-position distinctiveness as practical factors for fingerprint matching.

  • Impact of the Number of Packets per Batch: 10 packets per batch balances lower computation time with higher localization results, so DeepFi adopts this setting.Increasing batch size from 1 to 10 rapidly decreases average execution time; further increases produce only small changes, while larger batches reduce opportunities to average errors.
  • Impact of the Number of Packets per Batch: Batch processing keeps localization-error means and standard deviations approximately unchanged across six experiments.The reported stability is attributed to parallel processing that averages the errors of 100 packets.
  • Propagation Environment: More than 80% of test points retain strong CSI correlation when obstacles are farther from the wireless router.The paper attributes stronger correlation at greater distances to reduced distortion of strong signals such as the LOS signal.
  • Propagation Environment: Human mobility weakens CSI correlation more near the LOS path than near the reflected path, but CSI properties are not significantly affected overall.Fewer than 20% of LOS-path test points have correlation below 0.7, while approximately 20% of reflected-path test points fall below 0.8.
  • Training-Position Distinctiveness: Larger grid sizes reduce CSI similarity between neighboring positions, whereas smaller grids require redundant pre-training work.The paper frames grid selection as a trade-off between matching similar positions and limiting training overhead.

V. RELATED WORK

Related work spans fingerprinting, ranging, and angle-of-arrival approaches, with WiFi fingerprinting prominent because WiFi is ubiquitous indoors. DeepFi extends CSI-based fingerprinting with deep learning and is reported to outperform existing RSS- and CSI-based schemes in two indoor environments.

  • Overview: Indoor localization research is organized into fingerprinting-based, ranging-based, and angle-of-arrival categories.The supplied related-work passages describe these as the main categories of indoor localization systems.
  • Fingerprinting-based Localization: Fingerprinting-based localization surveys a floor plan during training and searches for the closest fingerprint during testing.WiFi fingerprinting is described as dominant because WiFi signals are ubiquitous indoors.
  • Fingerprinting-based Localization: Prior fingerprinting systems use diverse signals, including WiFi, FM radio, RFID, acoustic, GSM, light, and magnetism.WiFi-based systems include RSS methods such as RADAR and Horus and CSI methods such as FIFS and PinLoc.
  • Ranging-based Localization: Ranging-based methods estimate distances to at least three access points using power-based or time-based models.CSI-based ranging is presented as an approach intended to overcome RSS instability in indoor environments.
  • Angle-of-Arrival Localization: Angle-of-arrival systems estimate incoming angles with multiple antennas and then use geometric relationships to infer location.Synthetic-aperture approaches improve angular resolution but require accurate control of a moving antenna’s speed and trajectory.
  • DeepFi: DeepFi uses CSI in a deep-learning fingerprinting scheme and reportedly outperforms several existing RSS- and CSI-based schemes in two indoor environments.Its architecture analyzes CSI from all subcarriers and antennas with a four-hidden-layer network and uses trained weights as fingerprints.
Loading 1603.07080v1…