Smartphone-based pedestrian dead-reckoning (PDR) has become promising in indoor localization since it locates users with a smartphone only. However, existing PDR approaches are still facing the problem of accumulated localization errors due to low-cost noisy sensors and complicated human movements.ThispaperpresentsanovelPDRindoorlocalizationalgorithmcombinedwithonlinesequential extreme learning machine (OS-ELM). By analyzing the process of PDR localization, this paper first formulatestheprocessofPDRlocalizationasanapproximationfunction,andthen,asliding-window-based scheme is designed to preprocess the obtained inertial sensor data and thus to generate the feature dataset. At last, the OS-ELM-based PDR algorithm is proposed to address the localization problem of pedestrians. Due to the fact of universal approximation capability and extreme learning speed within OS-ELM, our algorithmcanadapttolocalizationenvironmentdynamicallyandreducethelocalizationerrorstoalowscale. Inaddition,bytakingthemovementhabitsofpedestrianintotheprocessofextremelearning,ouralgorithm can predict the position of pedestrian regardless of holding postures. To evaluate the performance of the proposed algorithm, this paper implements OS-ELM-based PDR on a real android-based smartphone and comparesitwiththestate-of-the-artapproaches.Extensiveexperimentresultsdemonstratetheeffectiveness of the proposed algorithm in various different postures and the practicability in indoor localization.
Pedestrian dead reckoning indoor localization based on os-elm
1. 1
Pedestrian Dead-Reckoning Indoor Localization
Based on OS-ELM
MINGYANG ZHANG, YINGYOU WEN, JIAN CHEN, XIAOTAO YANG,
RUI GAO AND HONG ZHAO
January 10, 2018
IEEE Access
2. 2
Motivation
The objective of this paper is to reduce the problems related to
accumulated localization error and complicated human
movements for indoor localization. A novel PDR indoor
localization algorithm combined with online sequential
extreme learning machine (OS-ELM) is used for localization.
3. 3
Contributions
Proposed the first OS-ELM based PDR algorithm.
Zero-crossing detection with a threshold-based peak detection for step detection.
The proposed system will not affect the different postures of holding the phone.
Designed a framework of OS-ELM based PDR for localizing pedestrians.
4. 4
Introduction to PDR
The pedestrian position can be
computed as
𝑥 𝑘+1
𝑦 𝑘+1
=
𝑥 𝑘
𝑦 𝑘
+𝑆𝐿 𝑘+1
sin(𝐻𝐷 𝑘+1)
cos(𝐻𝐷 𝑘+1)
(1)
The three procedures used in PDR can
be extracted as the following functions
𝑆𝐷 = 𝑓𝑠𝑑(𝑎)
𝐻𝐷 = 𝑓ℎ𝑑(𝑚, 𝑔)
𝑆L = 𝑓𝑠𝑙(𝑎)
Example of pedestrian dead-reckoning.
Where a, m and g are the values obtained from accelerometer, magnetometer and gyroscope.
𝑓ℎ𝑑 , 𝑓𝑠𝑑 and 𝑓𝑠𝑙 are the rules for estimating heading angles, detecting steps and estimating
stride length.
SD,HD and SL are the values of step detection, heading angles and stride length.
5. 5
STEP DETECTION
To overcome the tilting effect, the proposed algorithm transforms the raw acceleration
from smartphone coordinate system (SCS) to earth coordinate system (ECS).
To compute the acceleration in ECS, the proposed algorithm computes the rotation
matrix from SCS to ECS.
𝑅 𝑧 𝜓 𝑡 =
𝑐𝑜𝑠 𝜓 𝑡 𝑠𝑖𝑛 𝜓 𝑡 0
− sin 𝜓 𝑡 cos 𝜓 𝑡 0
0 0 1
𝑅 𝑥 𝜃𝑡 =
1 0 0
0 cos 𝜃𝑡 sin 𝜃𝑡
0 −𝑠𝑖𝑛𝜃𝑡 cos 𝜃𝑡
𝑅 𝑦 𝜙 𝑡 =
cos 𝜙 𝑡 0 sin 𝜙 𝑡
0 1 0
− sin 𝜙 𝑡 0 cos 𝜙 𝑡
Where 𝜓 𝑡, 𝜃𝑡 and 𝜙 𝑡 are the a azimuth angle, pitch angle and roll angle at the t-th sampling
moment.
The total rotation matrix of the z-x-y axes can be written as
𝑅𝑡
𝑧𝑥𝑦
= 𝑅 𝑧 𝜓 𝑡 𝑅 𝑥 𝜃𝑡 𝑅 𝑧 𝜙 𝑡 -------- (8)
Transformation of acceleration from SCS to ECS can be written as
𝑎 𝑡
𝐸𝐶𝑆
= 𝑅𝑡
𝑧𝑥𝑦
𝑎 𝑡
𝑆𝐶𝑆
------ ----- (9)
The z-axis component of the acceleration contains gravity, and then the proposed
algorithm eliminate the effect of gravity as
𝑎 𝑡
𝐿𝑖𝑛𝑒𝑎𝑟
= 𝑎 𝑡
𝐸𝐶𝑆
− 𝑔[0,0,1] 𝑇
6. 6
To reduce the effect of noise, the proposed algorithm performs a moving average filter
operation as
𝑎 𝑡 =
1
𝑚 𝑠𝑑
𝑖=𝑡−𝑚 𝑠𝑑+1
𝑡
𝑎 𝑧,𝑖
𝐿𝑖𝑛𝑒𝑎𝑟
Where the 𝑚 𝑠𝑑is the order of moving window. The filter linear acceleration 𝑎 𝑡 is used for
detecting steps. Example of step detection.
• This paper proposes an accurate step detection approach that combines the zero crossing
detection with peak detection.
7. 7
Stride Length and Heading Direction Estimation
The pedestrian stride length can be computed as
𝑆𝐿𝑖 = 𝑘
4
𝑎 𝑡𝑖
𝑃
− 𝑎 𝑡𝑖
𝑉
Where 𝑎 𝑡𝑖
𝑃
(𝑎 𝑡𝑖
𝑉
) is the peak (valley) of filtered linear acceleration at the i-th time step and K
is the coefficient.
𝑘 =
𝑖𝑆𝐿𝑖
4
𝑎 𝑡𝑖
𝑃
− 𝑎 𝑡𝑖
𝑉
𝑖𝑆𝐿𝑖
2
𝑎 𝑡𝑖
𝑃
− 𝑎 𝑡𝑖
𝑉
The heading angle at time t can be written as
𝐻𝐷𝑡 = 𝑓ℎ𝑑 𝑚 𝑡, 𝑔𝑡 = 𝜓 𝑡
The proposed algorithm replaces the aforementioned heading direction and stride length
estimation with an OS-ELM based localization approach.
9. 9
FRAMEWORK OF PROPOSED PDR LOCALIZATION
The framework contains two phases
1. The model training phase (dashed arrows)
Sensor data are processed into features and labels used for training OS-ELM models.
The proposed algorithm constructs two OS-ELM models The stride length estimation and heading
direction estimation
2. The PDR localization phase (solid-line arrows)
Estimates the stride length and heading direction by substituting the localization request data into trained
OS-ELM models.
12. 12
Specification
The threshold 𝛿 𝑎
+
and 𝛿 𝑎
−
to be 0.5
The size of sliding window W to be 20
Coefficient K to be 0.47
Moving average 𝑚 𝑠𝑑to be 3,
𝑚ℎ𝑑to be 15, 𝑚 𝑠𝑙 to be 4
Expanding times of heading direction
epochℎ𝑑
to be 5
Expanding times of stride length
epoch 𝑠𝑙to be 20
13. 13
SELECTION OF PARAMETERS FOR
OS-ELM MODELS
This paper evaluates the performance of three different activation functions:
1. Radial basis function
2. Sigmoid function
3. Sine function
• The number of hidden nodes for heading direction model : 300
• The number of hidden nodes for stride length model: 300
• The sine activation function is chosen as the activation function for stride length model
and heading direction model.
14. 14
EXPERIMENT RESULTS
The data in path 1 is chosen to compare the proposed step detection approach with some
popular step detection approaches. The relative error is employed to evaluate the
performance, which is defined as
𝑒 =
𝑁𝑒 − 𝑁𝑟
𝑁𝑟
× 100%
where 𝑁𝑒 is the number of detected steps, and 𝑁𝑟 is the ground truth.
15. 15
The performance of stride length and
heading direction estimation
To evaluate the performance of stride
length, the proposed approach is
compared with the typical linear
approach and nonlinear approach.
The path 2 is chosen to evaluate the
performance of heading direction estimation
approaches.
16. 16
Evaluation of the training time of the
proposed algorithm in real smartphone
In the experiment of training stride length model:
The total number of samples =1020
The total training time = 0.945
The training time of initialization phase =0.112s
The average training time of sequential phase = 0.0203s
In the experiment of training heading direction model:
The total number of samples =7105
The total training time = 40.35s
The training time of initialization phase =4.033s
The average training time of sequential phase = 0.1117s
• The training time of sequential learning phase can satisfy the requirement of online
learning.
• Therefore, it is practicable to deploy the propose localization algorithm in a real
smartphone.
17. 17
Conclusions
Proposed an OS-ELM based PDR indoor localization algorithm for android-based
smartphone.
The proposed localization algorithm does not force the smartphone to be held in
fixed posture.
Zero-crossing detection with a threshold based peak detection method is used
for step detection.
OS-ELM localization frame work is used for stride length and heading direction
estimation.
Sliding-window based scheme is used for preprocessing feature data.
The proposed PDR algorithm can continuously train OS-ELM online and generate
OS-ELM models for pedestrians movements.
The experiment results demonstrate the effectiveness of the proposed algorithm in
various different postures.