probabilistic modelling for sensor fusion (using …...probabilistic modelling for sensor fusion...

47
Probabilistic modelling for sensor fusion (using inertial sensors and magnetometers) Manon Kok Delft University of Technology, The Netherlands September 5th, 2018

Upload: others

Post on 15-Jul-2020

9 views

Category:

Documents


0 download

TRANSCRIPT

Page 1: Probabilistic modelling for sensor fusion (using …...Probabilistic modelling for sensor fusion (using inertial sensors and magnetometers) Manon Kok Delft University of Technology,

Probabilistic modelling for sensor fusion(using inertial sensors and magnetometers)

Manon Kok

Delft University of Technology, The Netherlands

September 5th, 2018

Page 2: Probabilistic modelling for sensor fusion (using …...Probabilistic modelling for sensor fusion (using inertial sensors and magnetometers) Manon Kok Delft University of Technology,

Brief background

• MSc in Philosophy of Science, Technology and Society at the Universityof Twente, the Netherlands.

• MSc in Applied Physics at the University of Twente, the Netherlands.

• Research Engineer at Xsens Technologies (February 2009 – July 2011).

• PhD in Automatic Control from Linkoping University (August 2011 –February 2017).

• Postdoc in the Machine Learning Group at the University of Cambridge(March 2017 – February 2018).

• Assistant Professor in the Delft Center for Systems and Control at theDelft University of Technology (April 2018 – Now).

2 / 29

Page 3: Probabilistic modelling for sensor fusion (using …...Probabilistic modelling for sensor fusion (using inertial sensors and magnetometers) Manon Kok Delft University of Technology,

Probabilistic modelling for sensor fusion(using inertial sensors and magnetometers)

3 / 29

Page 4: Probabilistic modelling for sensor fusion (using …...Probabilistic modelling for sensor fusion (using inertial sensors and magnetometers) Manon Kok Delft University of Technology,

About my researchSensor fusion: Using disparate sensors in order to infer more informationabout the quantity of interest than would be possible using each sensorindividually.

Quantity of interest: e.g. the position and orientation of the sensor.

• Inertial sensors• Accelerometers• Gyroscopes

• Magnetometers

4 / 29

Page 5: Probabilistic modelling for sensor fusion (using …...Probabilistic modelling for sensor fusion (using inertial sensors and magnetometers) Manon Kok Delft University of Technology,

Accelerometer measurements

5 10 15 20 25 30−20

0

20

Time [s]

y a,t

[m/s

2 ]

Stationary data ⇒ Earth gravityRotating sensor ⇒ Earth gravity

Quickly moving sensor ⇒ Earth gravity + sensor’s accelerationThe combination of earth gravity and sensor acceleration is called the

external specific force.

5 / 29

Page 6: Probabilistic modelling for sensor fusion (using …...Probabilistic modelling for sensor fusion (using inertial sensors and magnetometers) Manon Kok Delft University of Technology,

Accelerometer measurements

5 10 15 20 25 30−20

0

20

Time [s]

y a,t

[m/s

2 ]

Stationary data ⇒ Earth gravity

Rotating sensor ⇒ Earth gravityQuickly moving sensor ⇒ Earth gravity + sensor’s acceleration

The combination of earth gravity and sensor acceleration is called theexternal specific force.

5 / 29

Page 7: Probabilistic modelling for sensor fusion (using …...Probabilistic modelling for sensor fusion (using inertial sensors and magnetometers) Manon Kok Delft University of Technology,

Accelerometer measurements

5 10 15 20 25 30−20

0

20

Time [s]

y a,t

[m/s

2 ]

Stationary data ⇒ Earth gravityRotating sensor ⇒ Earth gravity

Quickly moving sensor ⇒ Earth gravity + sensor’s accelerationThe combination of earth gravity and sensor acceleration is called the

external specific force.

5 / 29

Page 8: Probabilistic modelling for sensor fusion (using …...Probabilistic modelling for sensor fusion (using inertial sensors and magnetometers) Manon Kok Delft University of Technology,

Accelerometer measurements

5 10 15 20 25 30−20

0

20

Time [s]

y a,t

[m/s

2 ]

Stationary data ⇒ Earth gravityRotating sensor ⇒ Earth gravity

Quickly moving sensor ⇒ Earth gravity + sensor’s accelerationThe combination of earth gravity and sensor acceleration is called the

external specific force.

5 / 29

Page 9: Probabilistic modelling for sensor fusion (using …...Probabilistic modelling for sensor fusion (using inertial sensors and magnetometers) Manon Kok Delft University of Technology,

Gyroscope measurements

2 4 6 8 10 12 14 16 18 20 22

−5

0

Time [s]

y ω,t

[rad

/s]

Stationary data ⇒ Zero angular velocityRotating the sensor around the y -axisRotating the sensor around the x-axisRotating the sensor around the z-axis

6 / 29

Page 10: Probabilistic modelling for sensor fusion (using …...Probabilistic modelling for sensor fusion (using inertial sensors and magnetometers) Manon Kok Delft University of Technology,

Gyroscope measurements

2 4 6 8 10 12 14 16 18 20 22

−5

0

Time [s]

y ω,t

[rad

/s]

Stationary data ⇒ Zero angular velocity

Rotating the sensor around the y -axisRotating the sensor around the x-axisRotating the sensor around the z-axis

6 / 29

Page 11: Probabilistic modelling for sensor fusion (using …...Probabilistic modelling for sensor fusion (using inertial sensors and magnetometers) Manon Kok Delft University of Technology,

Gyroscope measurements

2 4 6 8 10 12 14 16 18 20 22

−5

0

Time [s]

y ω,t

[rad

/s]

Stationary data ⇒ Zero angular velocityRotating the sensor around the y -axis

Rotating the sensor around the x-axisRotating the sensor around the z-axis

6 / 29

Page 12: Probabilistic modelling for sensor fusion (using …...Probabilistic modelling for sensor fusion (using inertial sensors and magnetometers) Manon Kok Delft University of Technology,

Gyroscope measurements

2 4 6 8 10 12 14 16 18 20 22

−5

0

Time [s]

y ω,t

[rad

/s]

Stationary data ⇒ Zero angular velocityRotating the sensor around the y -axisRotating the sensor around the x-axis

Rotating the sensor around the z-axis

6 / 29

Page 13: Probabilistic modelling for sensor fusion (using …...Probabilistic modelling for sensor fusion (using inertial sensors and magnetometers) Manon Kok Delft University of Technology,

Gyroscope measurements

2 4 6 8 10 12 14 16 18 20 22

−5

0

Time [s]

y ω,t

[rad

/s]

Stationary data ⇒ Zero angular velocityRotating the sensor around the y -axisRotating the sensor around the x-axisRotating the sensor around the z-axis

6 / 29

Page 14: Probabilistic modelling for sensor fusion (using …...Probabilistic modelling for sensor fusion (using inertial sensors and magnetometers) Manon Kok Delft University of Technology,

Inertial sensors for pose estimation

rotateremove

gravity

x

angular

velocity orientation

external spe-

cific force acceleration position

Accelerometers and gyroscopes (inertial sensors) can be used forposition and orientation (pose) estimation ⋯

7 / 29

Page 15: Probabilistic modelling for sensor fusion (using …...Probabilistic modelling for sensor fusion (using inertial sensors and magnetometers) Manon Kok Delft University of Technology,

Integration drift

however, the position and orientation estimates obtained using inertialsensors suffer from integration drift.

0 2 4 6 8 100

1

2

3

Time [s]

Orie

ntat

ion

[◦ ]

0 2 4 6 8 10

0

5

Time [s]

Posit

ion

[m]

⇒ Combine inertial sensors with other sensors and / or additionalmodels to obtain accurate estimates.

8 / 29

Page 16: Probabilistic modelling for sensor fusion (using …...Probabilistic modelling for sensor fusion (using inertial sensors and magnetometers) Manon Kok Delft University of Technology,

Probabilistic modelling for sensor fusion(using inertial sensors and magnetometers)

9 / 29

Page 17: Probabilistic modelling for sensor fusion (using …...Probabilistic modelling for sensor fusion (using inertial sensors and magnetometers) Manon Kok Delft University of Technology,

Probabilistic modelling

Probabilistic modelling to

• take into account that all models are approximations,

• take into account that all sensor measurements are not perfect,

• represent the best estimate and its uncertainty.

10 / 29

Page 18: Probabilistic modelling for sensor fusion (using …...Probabilistic modelling for sensor fusion (using inertial sensors and magnetometers) Manon Kok Delft University of Technology,

Probabilistic modelling for sensor fusion(using inertial sensors and magnetometers)

11 / 29

Page 19: Probabilistic modelling for sensor fusion (using …...Probabilistic modelling for sensor fusion (using inertial sensors and magnetometers) Manon Kok Delft University of Technology,

Magnetometers

Magnetometers measure the local(earth) magnetic field

12 / 29

Page 20: Probabilistic modelling for sensor fusion (using …...Probabilistic modelling for sensor fusion (using inertial sensors and magnetometers) Manon Kok Delft University of Technology,

Magnetometer measurements

2 4 6 8 10 12 14 16

−200

0

200

400

Time [s]

y m,t

[µT]

Stationary magnetometer data ⇒ Local (earth) magnetic fieldMagnetic disturbance close to sensor ⇒ Local (earth) magnetic field +

magnetic disturbance

Inertial sensors and magnetometers can be combined to estimateorientation.

But how to model the magnetometer measurements?

13 / 29

Page 21: Probabilistic modelling for sensor fusion (using …...Probabilistic modelling for sensor fusion (using inertial sensors and magnetometers) Manon Kok Delft University of Technology,

Magnetometer measurements

2 4 6 8 10 12 14 16

−200

0

200

400

Time [s]

y m,t

[µT]

Stationary magnetometer data ⇒ Local (earth) magnetic field

Magnetic disturbance close to sensor ⇒ Local (earth) magnetic field +magnetic disturbance

Inertial sensors and magnetometers can be combined to estimateorientation.

But how to model the magnetometer measurements?

13 / 29

Page 22: Probabilistic modelling for sensor fusion (using …...Probabilistic modelling for sensor fusion (using inertial sensors and magnetometers) Manon Kok Delft University of Technology,

Magnetometer measurements

2 4 6 8 10 12 14 16

−200

0

200

400

Time [s]

y m,t

[µT]

Stationary magnetometer data ⇒ Local (earth) magnetic fieldMagnetic disturbance close to sensor ⇒ Local (earth) magnetic field +

magnetic disturbance

Inertial sensors and magnetometers can be combined to estimateorientation.

But how to model the magnetometer measurements?

13 / 29

Page 23: Probabilistic modelling for sensor fusion (using …...Probabilistic modelling for sensor fusion (using inertial sensors and magnetometers) Manon Kok Delft University of Technology,

Magnetometer measurements

2 4 6 8 10 12 14 16

−200

0

200

400

Time [s]

y m,t

[µT]

Stationary magnetometer data ⇒ Local (earth) magnetic fieldMagnetic disturbance close to sensor ⇒ Local (earth) magnetic field +

magnetic disturbance

Inertial sensors and magnetometers can be combined to estimateorientation.

But how to model the magnetometer measurements?

13 / 29

Page 24: Probabilistic modelling for sensor fusion (using …...Probabilistic modelling for sensor fusion (using inertial sensors and magnetometers) Manon Kok Delft University of Technology,

My research

Goal:

Push the boundaries of accuracy that can be obtained from sensormeasurements by using

• advanced models,

• advanced algorithms.

Motivation:

• Sensors are becoming smaller, cheaper and more widely available.

• More and more computational power is available.

• More and more application areas are opening up.

14 / 29

Page 25: Probabilistic modelling for sensor fusion (using …...Probabilistic modelling for sensor fusion (using inertial sensors and magnetometers) Manon Kok Delft University of Technology,

Applications

Analysis of motion of:

• Cars

• Airplanes

• Quadcopters

• Trains⋮

Figures courtesy of Xsens Technologies

15 / 29

Page 26: Probabilistic modelling for sensor fusion (using …...Probabilistic modelling for sensor fusion (using inertial sensors and magnetometers) Manon Kok Delft University of Technology,

Applications

Figures courtesy of Xsens Technologies

Analysis of human motion for:

• Sports

• Rehabilitation

• Movies / games

• Robotics⋮

15 / 29

Page 27: Probabilistic modelling for sensor fusion (using …...Probabilistic modelling for sensor fusion (using inertial sensors and magnetometers) Manon Kok Delft University of Technology,

Two examples

• Human motion capture• Take as much information about the situation at hand into account.

• Indoor localisation using measurements of the magnetic field.• Extract as much information from sensor measurements as possible.

16 / 29

Page 28: Probabilistic modelling for sensor fusion (using …...Probabilistic modelling for sensor fusion (using inertial sensors and magnetometers) Manon Kok Delft University of Technology,

Example: Human motion capture

Estimate the relative position andorientation of body segments.

17 sensors placed on the body

Figures courtesy of Xsens Technologies

17 / 29

Page 29: Probabilistic modelling for sensor fusion (using …...Probabilistic modelling for sensor fusion (using inertial sensors and magnetometers) Manon Kok Delft University of Technology,

Example: Human motion capture

Estimate the relative position andorientation of body segments.

17 sensors placed on the bodyFigures courtesy of Xsens Technologies

17 / 29

Page 30: Probabilistic modelling for sensor fusion (using …...Probabilistic modelling for sensor fusion (using inertial sensors and magnetometers) Manon Kok Delft University of Technology,

Example: Human motion capture

Estimate the relative position andorientation of body segments.

17 sensors placed on the bodyFigures courtesy of Xsens Technologies

17 / 29

Page 31: Probabilistic modelling for sensor fusion (using …...Probabilistic modelling for sensor fusion (using inertial sensors and magnetometers) Manon Kok Delft University of Technology,

Model

Inertial sensors provide information aboutthe change in position and orientation.

The position and orientation of the sensorson the body is approximately constant.

The body segments are connected at thejoints. ⇒ constraint

Additional information...

18 / 29

Page 32: Probabilistic modelling for sensor fusion (using …...Probabilistic modelling for sensor fusion (using inertial sensors and magnetometers) Manon Kok Delft University of Technology,

Model

Inertial sensors provide information aboutthe change in position and orientation.

The position and orientation of the sensorson the body is approximately constant.

The body segments are connected at thejoints. ⇒ constraint

Additional information...

18 / 29

Page 33: Probabilistic modelling for sensor fusion (using …...Probabilistic modelling for sensor fusion (using inertial sensors and magnetometers) Manon Kok Delft University of Technology,

Model

Inertial sensors provide information aboutthe change in position and orientation.

The position and orientation of the sensorson the body is approximately constant.

The body segments are connected at thejoints.

⇒ constraint

Additional information...

18 / 29

Page 34: Probabilistic modelling for sensor fusion (using …...Probabilistic modelling for sensor fusion (using inertial sensors and magnetometers) Manon Kok Delft University of Technology,

Model

Inertial sensors provide information aboutthe change in position and orientation.

The position and orientation of the sensorson the body is approximately constant.

The body segments are connected at thejoints. ⇒ constraint

Additional information...

18 / 29

Page 35: Probabilistic modelling for sensor fusion (using …...Probabilistic modelling for sensor fusion (using inertial sensors and magnetometers) Manon Kok Delft University of Technology,

Model

Inertial sensors provide information aboutthe change in position and orientation.

The position and orientation of the sensorson the body is approximately constant.

The body segments are connected at thejoints. ⇒ constraint

Additional information...

18 / 29

Page 36: Probabilistic modelling for sensor fusion (using …...Probabilistic modelling for sensor fusion (using inertial sensors and magnetometers) Manon Kok Delft University of Technology,

Example: human body motion capture

arg minx1∶NT

− NT∑t=2

NS∑i=1

log p(xSit ∣ xSit−1, ySit )´¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¸¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¶

dynamics of the state xSit

− NS∑i=1

log p(xSi1 ∣ ySi1 )´¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¸¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¶prior

− NT∑t=1

NS∑i=1

log p(xBj

t ∣ xSit )´¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¸¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¹¶

placement of sensor Si on body segment Bj

subject to c(x) = 0

19 / 29

Page 37: Probabilistic modelling for sensor fusion (using …...Probabilistic modelling for sensor fusion (using inertial sensors and magnetometers) Manon Kok Delft University of Technology,

Human motion capture

Use only sensor data

Courtesy of Xsens Technologies

20 / 29

Page 38: Probabilistic modelling for sensor fusion (using …...Probabilistic modelling for sensor fusion (using inertial sensors and magnetometers) Manon Kok Delft University of Technology,

Human motion capture

Use sensor data + a biomechanical model

Courtesy of Xsens Technologies

21 / 29

Page 39: Probabilistic modelling for sensor fusion (using …...Probabilistic modelling for sensor fusion (using inertial sensors and magnetometers) Manon Kok Delft University of Technology,

Human motion capture

Use sensor data + a biomechanical model + a world model

Courtesy of Xsens Technologies

22 / 29

Page 40: Probabilistic modelling for sensor fusion (using …...Probabilistic modelling for sensor fusion (using inertial sensors and magnetometers) Manon Kok Delft University of Technology,

Two examples

• Human motion capture• Take as much information about the situation at hand into account.

• Indoor localisation using measurements of the magnetic field.• Extract as much information from sensor measurements as possible.

23 / 29

Page 41: Probabilistic modelling for sensor fusion (using …...Probabilistic modelling for sensor fusion (using inertial sensors and magnetometers) Manon Kok Delft University of Technology,

Example: Indoor localisation

Indoor localisation

• to find your way in largeshopping malls, train stations,etc,

• for first responders toemergencies,

• ⋮

0 1 2 3 4 5 10 m

30

40

50

60

Mag

nitu

de(µ

T)

Which sensors to use?

• GPS does not work (well) indoors.

• Use sensors that are present in any modern smartphone.

⇒ Use measurements of the magnetic field as a source of positioninformation!

24 / 29

Page 42: Probabilistic modelling for sensor fusion (using …...Probabilistic modelling for sensor fusion (using inertial sensors and magnetometers) Manon Kok Delft University of Technology,

Example: Indoor localisation

Indoor localisation

• to find your way in largeshopping malls, train stations,etc,

• for first responders toemergencies,

• ⋮0 1 2 3 4 5 10 m

30

40

50

60

Mag

nitu

de(µ

T)

Which sensors to use?

• GPS does not work (well) indoors.

• Use sensors that are present in any modern smartphone.

⇒ Use measurements of the magnetic field as a source of positioninformation!

24 / 29

Page 43: Probabilistic modelling for sensor fusion (using …...Probabilistic modelling for sensor fusion (using inertial sensors and magnetometers) Manon Kok Delft University of Technology,

Modelling the magnetic field

• Use machine learning to learn the magnetic field from data.

• Gaussian process models: Learn a map of the magnetic field. Alsolearn where you are certain or uncertain about this map.

25 / 29

Page 44: Probabilistic modelling for sensor fusion (using …...Probabilistic modelling for sensor fusion (using inertial sensors and magnetometers) Manon Kok Delft University of Technology,

Building magnetic field maps

Build a map of the indoor magnetic field to use it for localisation.

https://www.youtube.com/watch?v=enlMiUqPVJo

26 / 29

Page 45: Probabilistic modelling for sensor fusion (using …...Probabilistic modelling for sensor fusion (using inertial sensors and magnetometers) Manon Kok Delft University of Technology,

Magnetic field SLAM

Simultaneous localization and mapping

https:

//www.youtube.com/watch?v=enlMiUqPVJo

27 / 29

Page 46: Probabilistic modelling for sensor fusion (using …...Probabilistic modelling for sensor fusion (using inertial sensors and magnetometers) Manon Kok Delft University of Technology,

Two examples

• Human motion capture• Take as much information about the situation at hand into account.

• Indoor localisation using measurements of the magnetic field.• Extract as much information from sensor measurements as possible.

28 / 29

Page 47: Probabilistic modelling for sensor fusion (using …...Probabilistic modelling for sensor fusion (using inertial sensors and magnetometers) Manon Kok Delft University of Technology,

Thank you for your attention!

Questions?

https://www.tudelft.nl/staff/m.kok-1/

29 / 29