CCS355 Neural Network & Deep Learning UNIT III notes and Question bank .pdf
Motion recognition based 3D pedestrian navigation system using smartphone
1. 1Kyungpook National University
Motion Recognition-Based 3D Pedestrian
Navigation System Using Smartphone
Beomju Shin, Chulki Kim, Jaehun Kim, Seok Lee, Changdon Kee, Hyoung Seok Kim, and Taikjin Lee
IEEE SENSORS JOURNAL,
VOL. 16, NO. 18, SEPTEMBER 15, 2016
2. 2Kyungpook National University
Contents
i. ABSTRACT
ii. INTRODUCTION
iii. SYSTEM DESCRIPTION
iv. MOTION RECOGNITION
v. POSITIONING ALGORITHM OF PDR
vi. HEIGHT ESTIMATION ALGORITHM
vii. EXPERIMENTAL RESULTS
viii. CONCLUSION
3. 3Kyungpook National University
ABSTRACT
In existing inertial measurement unit (IMU)-based pedestrian dead-reckoning (PDR)
systems, sensor axes are fixed, because the IMU is mounted on the shoes or helmet.
The sensor axes of a smartphone are changed according to the walking motion of the
user, because the smartphone is usually carried by hand or kept in the pocket.
The conventional PDR method cannot apply to the smartphone-based PDR system.
To overcome this limitation, the walking status is detected using a motion recognition
algorithm with sensor measurements from the smartphone.
The proposed system detects the height information of the pedestrian.
(barometric pressure sensor)
The 3D position is calculated.
Advantages: cost, accessibility, no additional peripheral devices
(smartphone contains necessary sensors, such as an accelerometer, magnetometer, gyroscope,
and barometric pressure sensor.)
The proposed system gives high positioning accuracy.
4. 4Kyungpook National University
INTRODUCTION
GPS can be used in outdoor positioning.
Indoor position problems: Cost and Accessibility
Usage of smartphone increased in the world.
Indoor positioning consist of two technological
approach
1. Use the existence of infrastructure
Pseudolite GPS , Ultra-Wideband (UWB) , Bluetooth , radio frequency identification (RFID), ZigBee ,
wireless sensor network , Wi-Fi positioning system (WPS)
Provides accurate positioning results.
Requires extensive time and high costs for installation and maintenance of the infrastructure.
2. The inertial sensor-based system
This system can be susceptible to the drift error that accumulates over time.
It is cheaper and simpler than the first approach.
IMU-based PDR system
More Accurate.
Easier to implement.
To estimate the position of the pedestrian, the step length and heading
at each detected step are calculated through the use of the sensor data
from the IMU.
This system is Effective for tracking firefighters or soldiers.
This system is Effective for pedestrian indoor navigation.
5. 5Kyungpook National University
Smartphone-based indoor navigation system
More convenient than the IMU-based system for pedestrians.
The pedestrian using the smartphone can access the indoor
navigation service simply by downloading the application.
The smartphone has various onboard sensors : accelerometer,
gyroscope, and magnetometer
GPS and Wi-Fi in the mobile phone improves the position accuracy.
3D pedestrian navigation system using a smartphone ( This Paper propose)
Identical PDR algorithm is applied to estimate the step length and heading in all
forms of motion.
In the case of using a smartphone, the pedestrian typically holds the
smartphone in his or her hand, thus the axes of the smartphone move
continuously when the pedestrian is walking.
This makes it difficult to estimate the heading via the Kalman filter (KF).
To solve this problem, this study applies motion recognition to the
navigation.
The 2D position of the pedestrian is estimated using the motion recognition-based PDR and the height o
f the pedestrian is estimated using the barometric pressure sensor.
6. 6Kyungpook National University
SYSTEM DESCRIPTION
The 3D indoor navigation is based on the motion
recognition using the smartphone. The procedure of
the proposed system is presented in Figure.
The motion of the pedestrian is detected by the ANN
then the detected motion is utilized for PDR
positioning.
Certain PDR algorithm is used to detect pattern of the
motion.
Step detection algorithm is used to detect estimate motion.
For the heading estimation, the Kalman filter is used for viewing
motion and the pattern recognition method is used for swinging or
running motions.
The 2D position of the pedestrian is estimated
through the motion recognition-based PDR.
The height of the pedestrian is estimated using a baro
metric pressure sensor.
The 2D position and the height are integrated to
determine the 3D position of the pedestrian, and the
information is displayed on the smartphone.
System diagram of indoor navigation system
7. 7Kyungpook National University
MOTION RECOGNITION
There are several classification algorithms are used.
Decision tree (DT)
Support vector machine (SVM)
ANN
SVM provides the best results Implementation is difficult learning process requires an extended amount
of time.
ANN is easier process higher level of performance.
Number of input nodes – Feature number
Number of output nodes – Motions of the model
Number of hidden nodes – Flexible
To obtain the weights of the ANN, the back-propagation algorithm is
used in the feed-forward ANN architecture.
Backpropagation algorithm Decrease the error.
The training starts with random initial weights, and each weight is
adjusted until the error converges to a certain tolerance.
Architecture of an ANN model
8. 8Kyungpook National University
𝑇ℎ𝑒 𝑖𝑛𝑝𝑢𝑡 𝑣𝑒𝑐𝑡𝑜𝑟
𝑋𝑡 = {𝑥0, 𝑥1 … … . 𝑥𝑖}
t time , i size of input vector
𝑇ℎ𝑒 𝑜𝑢𝑡𝑝𝑢𝑡 𝑣𝑒𝑐𝑡𝑜𝑟
𝑌𝑡 = {𝑦0, 𝑦1 … … . 𝑦 𝑘}
k size of output vector
𝐼𝑛𝑝𝑢𝑡 𝑤𝑒𝑖𝑔ℎ𝑡𝑠
ℎ_𝑠𝑢𝑚 𝑗 =
𝑖=1
6
𝑤𝑖,𝑗, 𝑥𝑖
𝑊 input weight , 𝑥 element of input vector
B𝑖𝑝𝑜𝑙𝑎𝑟 𝑚𝑜𝑑𝑒 𝑎𝑐𝑡𝑖𝑣𝑎𝑡𝑖𝑜𝑛 𝑓𝑢𝑛𝑐𝑡𝑖𝑜𝑛
ℎ𝑗 = 𝜏 ℎ 𝑠𝑢𝑚 𝑗
𝜏 𝑥 =
2
1 + 𝑒−𝑎𝑥
− 1
𝑦_𝑠𝑢𝑚 𝑘 = 𝑗=1
6
𝑤𝑖,𝑘ℎ𝑗 𝑦 𝑘 = 𝜏(𝑦𝑠𝑢𝑚 𝑘
)
α slope parameter slope of the activation function.
α increases , the slope of the activation function becomes steep like a step function.
9. 9Kyungpook National University
Feature values in whole motion. Each feature value is changed
according to changes in motion of pedestrian.
Defined pedestrian motions. (a) M_1, (b) M_2, (c) M_3,
(d) M_4, (e) M_5, (f) M_6.
11. 11Kyungpook National University
POSITIONING ALGORITHM OF PDR
The PDR system consists of three parts:
1. Step Detection
2. Step length estimation
3. Heading estimation
Step Detection
To detect the step of the pedestrian, accelerometer is used and which gives three-axis acceleration values.
The root mean square (RMS) value is taken for the accelerometer reading in all three axes values as:
𝑎 𝑟𝑚𝑠 = 𝑎 𝑥
2
+ 𝑎 𝑦
2
+ 𝑎 𝑧
2 where 𝑎 𝑥 , 𝑎 𝑦 , and 𝑎 𝑧 indicate the acceleration values of each axis.
Three methods:
1. Zero crossing
2. Flat zone detection
3. Peak detection
zero crossing method detects the step when the signal passes a specific point detect a false
step for stationary motion because of an irregular behavior of the pedestrian.
The flat zone detection method is suitable for the foot-mounted IMU-based PDR.
The peak detection method that detects the highest peak in a sinusoidal wave.
12. 12Kyungpook National University
Without the averaging window
Multiple peaks appear in one period.
To avoid false peak detection, the moving
averaging filter is applied to the RMS value
of the accelerometer signal.
With the averaging window
The RMS value of the accelerometer reading after
applying the moving average filter.
RMS signal of the accelerometer
13. 13Kyungpook National University
RMS value of accelerometer
(a) RMS value at M_3. (b) RMS value at M_4. (c) RMS value at M_5.
(d) RMS value at M_6.
14. 14Kyungpook National University
Step length estimation
Once the step occurrence is detected, the length of the detected step is estimated.
a complex procedure.
Linear combination is used.
Linear combination consists of following parameters.
1. Acceleration Variance (AV)
2. Step Frequency (SF)
3. Gyroscope Integral (GI) value
These parameters are linearly proportional to step length.
𝑇ℎ𝑒 𝑎𝑐𝑐𝑒𝑙𝑒𝑟𝑎𝑡𝑖𝑜𝑛 𝑣𝑎𝑟𝑖𝑎𝑛𝑐𝑒
𝐴𝑉𝑘 =
1
𝑛
𝑖=𝑡 𝑘−1
𝑡 𝑘
(𝑎 𝑚 − 𝑎𝑖)2
n the number of sample data in one step.
𝑎 𝑚 the mean of the acceleration during one step.
𝑎𝑖 the accelerometer signal.
Relation between step length and parameters that
increase according to step length
15. 15Kyungpook National University
𝑆𝑡𝑒𝑝 𝐹𝑟𝑒𝑞𝑢𝑒𝑛𝑐𝑦 (𝑆𝐹)
𝑊𝐹𝑘 =
1
𝑛 × ∆𝑡
n the number of sample data during one step.
∆𝑡 the sampling rate of the sensor.
G𝑦𝑟𝑜𝑠𝑐𝑜𝑝𝑒 𝐼𝑛𝑡𝑒𝑔𝑟𝑎𝑙 𝐺𝐼 𝑣𝑎𝑙𝑢𝑒
𝐺𝐼 𝑘 =
𝑖=𝑡 𝑘−1
𝑡 𝑘
𝑔_𝑦𝑖 × ∆𝑡
𝑔_𝑦𝑖 is the y-axis signal of the gyroscope.
∆𝑡 sampling rate of the sensor
The step length of the pedestrian is estimated as follows:
𝑆𝐿 𝑘 =∝ ∙ 𝐴 𝑉𝑘 − 𝐴𝑉0 + 𝛽 ∙ 𝑊𝐹 𝐾 − 𝑊𝐹0 + ∙ 𝛾 ∙ 𝐺𝐼 𝐾 − 𝐺𝐼0 + 𝑆𝐿0
𝑆𝐿 𝑘 Estimated step length
𝑆𝐿0 Nominal step length
𝐴𝑉0, 𝑊𝐹0, 𝐺𝐼0 are the nominal values of AV, SF, and GI respectively.
16. 16Kyungpook National University
Heading estimation
The heading of the pedestrian is the most significant factor in the PDR.
The magnetometer and gyroscope are used in the heading estimation.
The Kalman filter (KF) is a useful tool when estimating the heading.
The gyroscope is used to update the time of the KF,
The magnetometer is used to update the measurements.
When the pedestrian performs M_2 (walking while looking at the phone)
The y-axis of smartphone is aligned with the direction of the pedestrian.
The z-axis becomes vertical to the ground.
In this case, the KF is used to conduct the heading estimation.
For other motions except for M_2,
The axes of the smartphone are changed continually during the walking motion.
The estimated heading is varied even when walking straight.
To solve this problem,
The pattern recognition is applied to the heading estimation for M_3 through M_6.
Motive of this approach
when the pedestrian moves in a particular direction while performing a particular motion, a magnetic field
from the magnetometer records a certain pattern, which is related to the present direction and motion.
Magnetic field is unique for each direction.
The heading could be estimated by classifying the different patterns of the magnetic field.
17. 17Kyungpook National University
The framework of the heading estimation
Feature values of the heading classifier when the pedestrian
is performing M_4 (walking while talking on the device).
18. 18Kyungpook National University
HEIGHT ESTIMATION ALGORITHM
Barometric pressure sensor can be used to estimate the height information of the pedestrian.
Barometric pressure sensor enables to distinguish which floor the user is located in a building.
Contributions of the proposed approach
Estimates the absolute height of the pedestrian.
The bias of each smartphone’s barometric pressure sensor is calculated using a sea-level pressure from the
Meteorological Office and a true altitude from the GPS.
Independent of time.
The thin lines in Figure represent the output of the
barometric pressure sensor received from two
smartphones that are in the same position and
time frame.
The output of the barometric pressure sensor shows
different results under the same conditions.
Reason Bias from each sensor.
To estimate an accurate height, the barometric pressure sensor needs to be calibrated.
Bold lines in Figure represent the calibrated output of the barometric pressure sensors from the two
smartphones.
19. 19Kyungpook National University
To calibrate the barometric pressure sensor, this study employs online sea-level pressure
information and a true altitude from the GPS in an outdoor environment.
𝐴 𝑟𝑒𝑓𝑒𝑟𝑒𝑛𝑐𝑒 𝑏𝑎𝑟𝑜𝑚𝑒𝑡𝑟𝑖𝑐 𝑝𝑟𝑒𝑠𝑠𝑢𝑟𝑒 𝑖𝑠 𝑜𝑏𝑡𝑎𝑖𝑛𝑒𝑑 𝑎𝑠 ∶ 𝑃𝑟(𝑡) = 𝑃0 𝑡 × 1 −
𝐻 𝑡
𝑆 𝑎
𝑆 𝑏
𝑃𝑟(𝑡) The reference barometric pressure
𝑃0 𝑡 The sea-level pressure
𝐻 𝑡 Altitude from the GPS at time t.
𝑆 𝑎, 𝑆 𝑏 scale factors.
The bias of the barometric pressure sensor is calculated as follows : 𝑃𝑏= 𝑃𝑟 𝑡 − 𝑃𝑚(𝑡)
𝑃𝑏 Bias
𝑃𝑚(𝑡) Measurement of the barometric pressure sensor at time t.
𝑇ℎ𝑒 𝑐𝑎𝑙𝑖𝑏𝑟𝑎𝑡𝑒𝑑 𝑜𝑢𝑡𝑝𝑢𝑡 𝑖𝑠 𝑜𝑏𝑡𝑎𝑖𝑛𝑒𝑑 𝑎𝑠: 𝑃𝑐 (𝑡) = 𝑃𝑚 𝑡 + 𝑃𝑏
𝑃𝑐 𝑡 The calibrated output of the barometric pressure sensor.
𝑇ℎ𝑒 ℎ𝑒𝑖𝑔ℎ𝑡 𝑖𝑠 𝑜𝑏𝑡𝑎𝑖𝑛𝑒𝑑 𝑎𝑠 𝑓𝑜𝑙𝑙𝑜𝑤𝑠: 𝐻 𝑚 𝑡 = 𝑆 𝑎 × (1 − (
𝑃𝑐 𝑡
𝑃0 𝑡
)
1
𝑆 𝑏)
The present floor is obtained using the estimated height as follows:
𝐹 𝑡 =
𝐻 𝑚 𝑡 − 𝐻0 𝑡
𝐷
𝐹 𝑡 Estimated present floor
𝐻0 𝑡 1st floor height of the building
D Interfloor height
Calibration
20. 20Kyungpook National University
EXPERIMENTAL RESULTS
Motion Recognition Result
Result of motion classifier
The ANN-based motion classifier provides a very
high level of performance for all motions.
The motion classifier predicts the motion perfectly
except during the transition of motion.
Step Length Estimation Result
Results of step length estimation
The pedestrian walked 144 m in various
walking speeds, such as slow, normal, and fast
per 48 m.
The RMSE of the step length estimation
is 2.34 cm and the error rate is 2.88%.
21. 21Kyungpook National University
Height Estimation Result
To obtain the barometric pressure data, the pedestrian went from the 1st floor up to the
rooftop and down to the 1st basement floor carrying the two calibrated smartphones.
It is noticeable from the table that the false floor estimations before calibration are corrected
to yield the true floor information after calibration.
The result of calibrated barometric pressure sensors
22. 22Kyungpook National University
CONCLUSION
This paper proposed the motion recognition-based 3D pedestrian navigation system using smartphones.
The proposed PDR system estimates the position regardless of any motion performed by the pedestrian.
The ANN-based motion classifier is employed to recognize the motion.
The height of the pedestrian is estimated using the onboard barometric pressure sensor of the
smartphone.
The barometric pressure sensors are calibrated using the sea-level pressure from the Meteorological
Office and altitude from the GPS.
The proposed system provides positioning results that consist of an error rate of less than 6%.
FUTURE WORKS
Creating heading classifiers for eight directions.
Developing the motion and heading classifier applicable to various persons.
Enhancing the resolution of the heading classifier.