[time]
2020-11-19T12:10:09+01:00

[team_name]
WiMaP

[team_institution]
Beijing University of Posts and Telecommunications; Institute of Computing Technology, Chinese Academy of Sciences

[logolink]


[system_name]


[website]


[track]
3

[reference_person]
Qu Wang

[email]
wangqu@ict.ac.cn

[description]
ABSTRACT
Smartphones have become an indispensable tool in the work and life of most people. Given the many kinds of sensors built into smartphones, solving indoor navigation problems using these phones has become feasible. This paper proposes such a solution, using a neural network-based indoor positioning method. This method improves the positioning accuracy and helps improve the user experience. The major features of the method are that it leverages and fuses the advantages of two different indoor positioning methods [pedestrian dead reckoning (PDR) and neural network-based Wi-Fi indoor positioning], and using particle filter to considers the influence on positioning of the indoor spatial map. In the training process, the smartphone built-in sensors are used as data sources to provide fine-grained position observations. The sparse Wi-Fi signal are trained for regressing a position which label is provide by PDR mothed. In the test process, Wi-Fi signal are used as observations in the particle filter. 
Keywords: Indoor localization; smartphone; deep learning.
 1.   INTRODUCTION
Indoor positioning has attracted great attention and has been growing rapidly in recent years. Navigation and common daily services such as restaurant search and marketing are typical location-based service applications, which rely on the (easy to obtain) user’s real-time location. Global navigation satellite systems especially the global positioning system (GPS) have been widely used to provide an accurate localization service in outdoor environments. According to the recent market report, ordinary people usually spend more than an average of 70% of their time indoors. In this case, indoor positioning is of vital importance to human live; however, GPS does not work properly indoors because of signal attenuation. Therefore, a few techniques and devices have been proposed and used for indoor positioning: wireless fidelity (Wi-Fi), Blue-tooth low energy (BLE), ZigBee, radio frequency identification (RFID), ultra-wideband, etc. Among these, Wi-Fi and BLE are the most used. Wi-Fi or iBeacon (a BLE implementation by Apple) devices usually act as access points (APs), so that the received signal strength indicator (RSSI) can be used to achieve indoor positioning.
2.   SYSTEM DESIGN
The whole system combines data pre-processing, Pedestrian Dead Reckoning, Wi-Fi fingerprint training, particle filter, as shown in Fig 1.
 
Fig. 1. Particle filter with the map constraint scheme, using Wi-Fi RSSI results and IMU data as inputs.
2.1 Pedestrian Dead Reckoning
Most smartphones have built-in sensors that measure motion, orientation, and various other environmental conditions. These sensors are capable of providing raw data with high precision and accuracy and are particularly useful if we want to monitor 3-D device movements or positioning (such as tilt, vibration, rotation, and rocking), or use them to monitor changes in the ambient environment near a device (temperature and pressure) . With the data from IMU sensors, parameters such as step frequency, step length, and heading direction can be calculated. PDR-based indoor positioning can be expressed by the following equations:
x_k=[█(e_k@n_k )]=[█(e_(k-1) l_k·sin(θ_k )@n_(k-1) l_k·cos(θ_k ) )] [█(δ_e@δ_n )]	(1)
δ^2=δ_e^2 δ_n^2	(2)
where l_k is the distance traveled on step k, θ_k is the corre- sponding heading direction, e_(k-1) and n_(k-1) represent the east and north coordinates before step k,and e_k and n_k denote the east and north coordinates after step k. δ_eand δ_n represent the errors in the east and north directions. δ_e and δ_n are introduced here to match with the observed uncertainty measurement in the step length and heading direction estimated from the IMU measurements.
2.2 Wi-Fi Fingerprint Training
The Wi-Fi hotspots provide solid location information by wireless signal. Each Wi-Fi access points has an independent SSID and MAC address to identify accurate locations. The cluster of Wi-Fi access points’ signal features constitutes the relations to an accurate location. The technique of Wi-Fi consists in 4 main types, such as received signal strength indication (RSSI), fingerprints, angle of arrival (AoA) and time of flight (ToF) based techniques. The provided logs in the challenge only contains RSSI of each hotspot, so we use the RSSI cluster to build the fingerprint of accurate location.
We provide a deep learning-based method to regress the position. Counting the number of hotspots, we listed all of Wi-Fi hotspots for a total of 450 and fill the RSSI of each hotspot by replacing the missing value to -120Hz. By regressing using deep learning model, a related position value can be produced to join the reference value list of absolute positions, which can be optimized in following stages. The deep learning model is displayed in figure 2. In the regressing model, we use a convolutional layer to reduce the sparseness of the RSSI vector. The feature extractor layers constructed by a ResNet are fed by the extracted dense vector of features. 
 
Fig. 2. The deep learning pipeline of Wi-Fi RSSI fingerprint location
2.3 Particle Filter
The key strategy of particle filters is to approximate the posterior probability distribution by many random experiments and then replace integral operations by a sample mean, thus obtaining the minimum variance distribution. They can be used with any form of state-space models and have been described in detail in [20] and [30]. The positioning process is constituted by the following steps. 
	Initialization: Generate an initial set of N state vectors—referred to as N particles—from the initial probability density and assign them equal weights.
{X_0^i }   ,ω_0^i  = 1/N.	(3)
	Prediction: Determine a new position for each particle using PDR [referring to (1)] and a random noise realization
X_k^i=[█(e_(k-1)^i l_k·sin(θ_k )@n_(k-1)^i l_k·cos(θ_k ) )] [█(δ_e^i@δ_n^i )]	(4)
where l_k and θ_kcan be calculated from the above sections; δ_e^i and δ_n^i are random noise terms, normally distributed, with zero mean and given variances. The variances can be set to be the root-mean-square errors (RMSEs) in different directions obtained from PDR experiments.
	Update: The observation equation is based on RSSI positioning and its errors.
Z_k=[█(e_k@n_k )] [█(σ_e@σ_n )]	(5)
σ^2=σ_e^2 σ_n^2	(6)
where e_k and n_k are the positioning result; the noise terms σ_e and σ_n can be represented by the RMSE of RSSI in different directions obtained from experiments.
The Gaussian pseudo distribution is often used to update the weights, as presented in the following equation:
ω_k^i=α_k^i·β_k^i·1/(√2π σ)·e^(-((X_k^i-Z_k )^T·(X_k^i-Z_k ))/(2σ^2 ))	(7)
where α_k^i indicate whether a particle is in an accessible area. When a particle is in inaccessible area (α_k^i = 0), it will be treated as an invalid particle. Doors are the only way to reach accessible areas; if a particle moves from one accessible area to another accessible area without passing through the door, it is considered as having passed through a wall, and its particle weight will, therefore, be set to 0, which means that β_k^i= 0. After calculating all particle weights, we must normalize them
¯ω_k^i=(ω_k^i)/((∑ω_k^j)┬(j=1))┴N 	(8)
	Resample: In this step, importance resampling [20] is used to obtain a new particle set {X_k^i }.
	State Estimation: Estimate the state by
X_k=1/N  ( (∑X_k^i)┬(i=1))┴N	(9)
	Iterate: Set k = k   1 and go to step 2.


[references]
[1]	R. Wang, H. Luo, Q. Wang, Z. Li, F. Zhao, and J. Huang, “A Spatial-Temporal Positioning Algorithm Using Residual Network and LSTM,” IEEE Trans. Instrum. Meas., vol. 69, no. 11, pp. 9251–9261, 2020.
[2]	Q. Wang et al., “Pedestrian Dead Reckoning based on Walking Pattern Recognition and Online Magnetic Fingerprint Trajectory Calibration,” IEEE Internet Things J., vol. 4662, no. c, pp. 1–1, 2020.
[3]	Q. Wang et al., “Personalized Stride-Length Estimation Based on Active Online Learning,” IEEE Internet Things J., vol. 7, no. 6, pp. 4885–4897, Jun. 2020.
[4]	R. Wang, Z. Li, H. Luo, F. Zhao, W. Shao, and Q. Wang, “A Robust Wi-Fi Fingerprint Positioning Algorithm Using Stacked Denoising Autoencoder and Multi-Layer Perceptron,” Remote Sens., vol. 11, no. 11, p. 1293, May 2019.
[5]	Q. Wang, H. Luo, A. Men, F. Zhao, and Y. Huang, “An Infrastructure-Free Indoor Localization Algorithm for Smartphones,” Sensors, vol. 18, no. 10, p. 3317, Oct. 2018.
[6]	Q. Wang, L. Ye, H. Luo, A. Men, F. Zhao, and C. Ou, “Pedestrian Walking Distance Estimation Based on Smartphone Mode Recognition,” Remote Sens., vol. 11, no. 9, p. 1140, May 2019.