Robust pedestrian dead reckoning (R-PDR) for arbitrary

2014 International Conference on Indoor Positioning and Indoor Navigation, 27th -30th October 2014
Robust pedestrian dead reckoning (R-PDR) for
arbitrary mobile device placement
Zhuoling Xiao, Hongkai Wen, Andrew Markham, and Niki Trigoni
Department of Computer Science, University of Oxford.
Email: [email protected]
Abstract—Pedestrian dead reckoning, especially on smartphones, is likely to play an increasingly important role in indoor
tracking and navigation, due to its low cost and ability to work
without any additional infrastructure. A challenge however, is
that positioning, both in terms of step detection and heading
estimation, must be accurate and reliable, even when the use of
the device is so varied in terms of placement (e.g. handheld or in
a pocket) or orientation (e.g holding the device in either portrait
or landscape mode). Furthermore, the placement can vary over
time as a user performs different tasks, such as making a call
or carrying the device in a bag. A second challenge is to be able
to distinguish between a true step and other periodic motion
such as swinging an arm or tapping when the placement and
orientation of the device is unknown. If this is not done correctly,
then the PDR system typically overestimates the number of steps
taken, leading to a significant long term error. We present a
fresh approach, robust PDR (R-PDR), based on exploiting how
bipedal motion impacts acquired sensor waveforms. Rather than
attempting to recognize different placements through sensor data,
we instead simply determine whether the motion of one or both
legs impact the measurements. In addition, we formulate a set of
techniques to accurately estimate the device orientation, which
allows us to very accurately (typically over 99%) reject false
positives. We demonstrate that regardless of device placement,
we are able to detect the number of steps taken with >99.4%
accuracy. R-PDR thus addresses the two main limitations facing
existing PDR techniques.
Keywords—Indoor positioning, inertial tracking, dead reckoning, heading tracking.
I.
I NTRODUCTION
Pedestrian Dead Reckoning (PDR) is widely viewed as
being a promising technique for enabling indoor positioning
and navigation with low power and cost, and without requiring
any infrastructure. Accurate and robust indoor localization is
critical to bring location-based and location-aware services
into widespread use. This is of ever growing importance
as emerging device classes like smartwatches and glasses
enter the wearable ecosystem and more and more applications
require the use of accurate positioning.
Pedestrian dead reckoning is conceptually a simple technique - a user’s displacement from a starting point is given by
the number of steps they have taken and what direction each
step was taken in. Typically, this is achieved with the aid of
an IMU, consisting of an accelerometer (linear acceleration),
gyroscope (rotational acceleration) and a magnetometer (to
determine the heading relative to magnetic North). This is in
contrast with inertial based techniques which double integrate
the raw sensor values to estimate a platform’s displacement
between sampling time instants, leading to rapid accumulation
of errors, resulting in excessive drift after short time periods.
PDR works particularly well when the sensor is rigidly attached to a person (i.e. a strapdown sensor) and mounted on
the foot, and have helped to demonstrate its potential.
However, for mobile devices (such as smartphones), the
requirement for rigid attachment and specified placement are
incompatible with the way in which people use devices. For
example, over a period of a few hours, a smartphone could be
carried in a backpack and then shifted to a pocket, before being
taken out and being used to send a text message. Each of these
activities are independent of whether or not the user is walking.
The PDR system must operate correctly and consistently in
all these scenarios. The challenge of accurately estimating
steps and heading without knowledge of the device placement
on the user is extremely challenging and a major hurdle
that is preventing widespread adoption of PDR on consumer
technology.
For the first component of PDR, step detection, an error
of even 2% over a long trajectory (e.g. 20 steps error in
1000 steps taken) can place the user in an incorrect room,
leading to the potential failure of context aware services. This
is because errors accumulate over a long trajectory, leading
to increased uncertainty over time. Ever more sophisticated
algorithms have been proposed to perform the task of motion
or activity recognition, with the idea that if the placement
can be inferred, then a case-specific step detection algorithm
can be executed. One major issue that we have noticed with
existing step detection techniques is that they are particularly
poor at distinguishing between rhythmic signals caused by
walking and signals caused by behaviours with strong periodic
components like typing, tapping a finger, nodding one’s head
and tapping a foot. For a PDR technique to be widely adopted,
it must be able to reject these false positives or else the
estimated position of the user rapidly deviates from the true
position.
The second component of a pedestrian dead reckoning
system is being able to accurately determine the heading
(forward direction of the user). Typically, this is achieved when
the user is stationary i.e. at the zero velocity point. This is
especially challenging if the device orientation is not known,
as a translation from sensor to body frame coordinates needs to
be made. Moreover, if the user changes the orientation of the
device e.g. rotating it in their hand from portrait to landscape
mode, tracking can fail for existing techniques.
We propose a fresh approach to PDR that takes a step back
from existing approaches and instead looks at the fundamentals
of human bipedal motion and how this would be registered at
various points of the human body. Based on this, we show that
human motion as detected by an arbitrarily placed device falls
into two categories, namely symmetric motion and asymmetric
2014 International Conference on Indoor Positioning and Indoor Navigation, 27th -30th October 2014
Fig. 1.
System architecture.
motion. Symmetric motion accounts for placements where both
footfalls are detected, e.g. if a user is making a call and walking
at the same time. Asymmetric motion, on the other hand,
captures the case when the dominant motion component is
registered on one side of the body only, e.g. if a user is walking
with the phone in their hand. This simple binary decision
allows us to build a very robust step detection system that does
not need a complex, case-specific motion mode recognition
algorithm. Similarly to address the issue of determining the
heading of the user when the orientation of the device is
unknown, we again turn to basic principles and attempt to
determine the principal swing axis of the user, rather than the
heading of the phone. We present a new technique for dead
reckoning, Robust-PDR (R-PDR).
In particular, the following are specific contributions of this
paper:
•
Robust PDR with arbitrary device placement: By
carefully examining every block of a PDR system,
we develop techniques that attain extremely high step
detection accuracy (>99.4%), regardless of where the
device is sited.
•
Rejection of false steps: We demonstrate excellent
rejection of non-walking human motion, again regardless of device placement.
•
Extensive real-world experiments: we show how we
can greatly increase the accuracy of PDR systems by
combining these suite of techniques in extensive real
world tests with unconstrained placement.
The remainder of this paper is organized as follows: Sec. II
outlines the system architecture of R-PDR, motivating the
need for each component. We first discuss how to accurately
estimate device orientation in Sec. III, key to accurate step
detection. Sec. IV presents the first component of the dead
reckoning system that accurately counts the number of steps
taken, regardless of device placement, whereas Sec. V illustrates how heading of the user can be accurately estimated
by examining the principal swing axis of the device. Sec. VI
extensively evaluates the proposed algorithm in a variety
of techniques and compares it with competing techniques.
Sec. VII contrasts our proposed work with existing techniques
and lastly Sec. VIII summarises the contributions of the paper.
II.
S YSTEM A RCHITECTURE
The block diagram of the system architecture is shown
in Fig. 1. In this work, we focus on building robust and
lightweight techniques for accurate pedestrian dead reckoning.
Like many PDR systems, it comprises three main components:
a fusion algorithm to convert raw sensor readings to estimates
of device motion; a step detector; and a heading estimator.
However, unlike existing PDR systems, it is intended for use
in arbitrary device placement and orientation. Each component
has thus been specifically tailored towards these challenges,
exploiting fundamental characteristics of human motion. In
addition, we have aimed to produce a lightweight system that
is applicable to multiple settings (smartphone, smartwatch etc.)
without requiring retuning or calibration. The core components
are discussed in detail below.
A. Inertial Fusion
This comprises two main blocks, a standard Kalman Filter
to combine data from the 3 classes of motion sensors, and an
orientation correction algorithm (RIOT) that constrains long
term orientation drift (especially in the pitch and roll axes)
which arises as a result of bias in gyroscope measurements.
This is made especially challenging as the device placement
can be arbitrary and dynamically changed. By using short term
estimates of relative rotation, estimates of gravity can be made
invariant to the rotation of the device over time.
B. Step detector
Candidate Step Detection: Whereas existing work has typically built more and more elaborate models to recognize
different classes of motion [22], [23], we instead take a
simpler, more robust view, based on fundamentals of bipedal
motion and how it couples to different device placements. The
periodicity detector assesses whether the sensed motion of the
user is symmetric or asymmetric. Informally, the symmetric
motion modes correspond to the class of device placements
that are impacted in a similar manner by both the right and left
leg (e.g. making a call whilst walking), whereas asymmetric
motion corresponds to placements where the movement of only
one leg impacts device motion (e.g. holding the device in a
hand and walking). In general, we find that symmetric motion
corresponds to strong periodicity in the acceleration domain,
whereas asymmetric motion corresponds to strong periodicity in the orientation domain. The step detector determines
2014 International Conference on Indoor Positioning and Indoor Navigation, 27th -30th October 2014
whether it is possible that a step has been taken, using an
enhanced zero crossing detector. It must be noted at this point
that the steps are treated as being candidates only, as it is
possible that another motion (like shaking or tapping) would
correspond to a similar looking acceleration pattern.
Repetitive Pattern Update (REPUT): By exploiting constraints imposed on human steps, long term orientation drift,
especially in the horizontal plane, can be greatly reduced.
This novel algorithm attempts to correct the orientation bias
introduced by the gyroscope.
Step Validation: Given potential step events, this block verifies whether a step really was taken, key to the accuracy of
our PDR system. It does this by estimating the displacement
made in both the horizontal and vertical directions by double
integrating the acceleration readings.
C. Heading Estimation
Using detected steps, the last task is to accurately determine
the direction of travel of the user. In the majority of existing
work, the heading is assumed to be parallel to one of the axes
of the device. However, even with simple, arbitrary placements
like texting, the heading of the device is not necessarily the
same as the heading of the user. To tackle this problem,
we project the displacement vectors of the step onto a two
dimensional plane and determine the principal axis of travel.
D. Positioning
In this final block, a mapping application is used to turn
dead reckoned steps and heading into user trajectories. We use
an existing state-of-the-art map matching algorithm, MapCraft
[24], which uses conditional random fields to constrain trajectories, to estimate positions. However, any technique such
as particle filters or extended Kalman filtering can be used to
determine position.
III.
ROTATIONALLY I NVARIANT O RIENTATION T RACKER
(RIOT)
Being able to accurately track the device’s orientation is
the foundation for accurate long term inertial tracking. This is
because most PDR subtasks, such as step detection and user
heading estimation, are based on knowing the orientation of
the device. However, determining device orientation remains
a major issue for inertial tracking with unconstrained devices,
especially with low-cost IMU sensors ubiquitous in mobile
devices and wearable sensors. Generally, the gyro drift (and to
a lesser degree, magnetic distortions from the environment) is
the major obstacle for accurate orientation tracking.
Most existing work uses Kalman filters to track device
orientation, with measurements of the Earth’s magnetic field
to correct the long term drift of the yaw angle. However, the
gyro drifts in roll and pitch angles remain an issue. We argue
that this problem can be addressed by accurately estimating the
gravitational acceleration from the raw accelerometer signal,
and using the gravity vector as an observation in the Kalman
filter to correct drift in both roll and pitch.
Currently there are two dominant approaches to estimating
the gravity component of accelerometer measurements. The
first approach, proposed by Mizell et al. [14], calculates
the mean over a fixed length window to remove dynamic
acceleration and estimate gravity as the static component.
This approach is elegant and simple but is not suitable for
mobile devices used by pedestrians. The major reason is
that a change in orientation introduces a considerable lag in
accurate gravity estimation. To reduce the lag, more recent
work [13] reduces the length of the window, with a consequent
decrease in accuracy. The second approach, recently proposed
by Hemminki et. al [9], places additional constraints over
allowable sampling instants for gravity estimation. Specifically,
this approach estimates the direction of the gravitational vector
only when the variation of the acceleration signals over a
window is below a certain threshold. This reduces the lag
introduced by the first approach, but also suffers from two
fundamental limitations. First, this approach does not always
lead to enough samples for estimating the gravitational vector,
as it only works when the mobile device is static for a period of
time. It typically does not work if the user is walking, when
accurate estimates of the gravity vector are mostly needed.
Secondly, the thresholds of variation in different motion modes
are different, making it impractical in real-world settings, as
significant tuning is required.
To address these limitations, we propose a novel algorithm
that can accurately estimate the gravity vector without lag and
operates continually. The key property that it exploits is that
gyro sensors have low drift and high reliability over a short
time window, and thus can provide accurate measurements of
relative changes in orientation. Assume we want to estimate the
gravity vector at time t given a window of historical accelerometer data at−T :t of size T . We first rotate acceleration signals
at−T :t−1 to the same orientation as at , and then estimate the
gravity at time t as the mean of the acceleration readings after
the rotation. In this way, we address the limitation of the first
approach in needing a long window for accurate measurements
of device orientation and avoids the constraint of the second
in needing the device to be stationary for a length of time.
Specifically, the gravity at time t is estimated as
gt = Rt
t
X
RTτ aτ
,
T
(1)
τ =t−T
where aτ is the acceleration vector in the coordinate frame of
the device, and Rτ is the rotation matrix describing the rotation
from the earth coordinate system to the coordinate system
of the device at time τ , as obtained from the Kalman Filter.
This value is then used to compensate the Kalman Filter, by
supplying an accurate observation of the gravity, thus removing
the effects of long term gyro drift. Typically, we find that a
window length of 4 seconds leads to good results.
Experiments have been conducted to compare the proposed algorithm with the state-of-the-art gravity estimation
techniques. In the experiment the pedestrian walks normally
with the mobile device carried in their hand. The gravity vector
is estimated using the proposed and competing approaches and
then rotated to the earth’s coordinate system according to the
current orientation of the device. The gravity value, as shown
in Fig. 2, is then estimated as the acceleration in the Z axis
(vertical axis). If the orientation algorithm is working correctly,
the gravity vector should only have a component in the vertical
axis, and this value should be equal to the true value of
gravity. It is observed that Mizell’s approach has the largest
errors due to its inability to manage changes in orientation
of the device. Meanwhile, Samuli’s approach provides very
Acceleration (m/s2)
−6
−8
Acceleration (m/s2)
−4
15
Acceleration (m/s2)
2014 International Conference on Indoor Positioning and Indoor Navigation, 27th -30th October 2014
15
−10
−12
−14
−16
10
15
20
Time (s)
Acceleration
Gravity
Mizell et. al
Samuli et. al
R−PDR
25
30
Fig. 2. The acceleration due to gravity estimated along the Z axis using
different techniques.
sparse and sometimes inaccurate gravity estimates, because in
most cases the motion of the pedestrian makes the variation
in acceleration bigger than the threshold. In comparison, the
proposed approach offers accurate gravity estimates and works
continually. Both the proposed approach and Mizell’s approach
use a window length of 4 s.
Fig. 3 shows the long term estimate of gravity with
and without the proposed orientation tracking algorithm in
the Z axis. Again, we would expect a good technique to
maintain the estimate of the gravity component at a constant
-9.81m/s2 , regardless of the orientation of the device or
dynamic acceleration changes due to motion. As expected, the
drift of device orientation grows quickly without the proposed
approach, leading to the gravity value drifting away from the
actual. This drift can result in inaccuracy in counting number
of steps, as will be explained in Sec. IV. However, the drift
has been successfully corrected by the proposed orientation
tracking approach, which has always kept the estimated gravity
value very close to its true value.
IV.
S TEP D ETECTION
In this section, we discuss our novel techniques for robust
step detection. Key is to detect true steps accurately, whilst
disregarding motion signatures such as shaking or tapping
that could be mistaken for steps. Whereas existing work has
typically built more and more elaborate models to recognize
different classes of motion [22], [23], we instead take a simpler,
more robust view, based on fundamentals of bipedal motion
and how it couples to different device placements. The step
detection component comprises three core blocks. The first
block examines the pitch and vertical acceleration waveforms
to determine whether a candidate step has been taken, based
on periodicity. The second block further refines estimates of
acceleration in the horizontal plane (forward direction), based
on repetitive features common to multiple steps. The third and
final block validates whether or not a step was really taken,
by examining the overall displacement of the device over the
duration of the step interval.
A. Candidate Step Extractor
The first block in the step detector extracts any potential
candidate steps from the motion measurements. An approach
10
5
Acceleration
Gravity
Estimated gravity
100
300
400
Time (s)
10
5
0
200
500
600
Fig. 3. The resulting acceleration signal in the Z axis without (top) and with
(bottom) the rotationally invariant orientation tracker.
for detecting steps from unconstrained device placement is
to build a classifier that recognizes where on the body the
device is placed [23], [25]. We take a fresh stance, based on
fundamentals of bipedal motion and resultant sensor coupling
for arbitrarily placed devices. Walking motion consists of two
phases, the stance phase (when the foot is in contact with
the ground) and the swing phase (when the foot is moving
forward in an arc), which each leg undertakes in a rhythmic,
offset pattern. When walking, the arms also typically swing
back and forth. The motion of the legs and arms couples
through the body, and depending on where the sensor is placed,
it will either pick up equal and symmetric forces from both
sides of the body, or asymmetric forces from only one half of
the body, with the other side of the body weakly coupling
to the sensor. An example of symmetric motion would be
making a call whilst walking, as the smartphone would be held
relatively rigidly against the head and thus would be impacted
by the movements of both legs equally. Conversely, holding
the device in one’s hand and swinging the arms would lead to
strongly asymmetric waveforms. We argue that by exploiting
these basic observations, the task of classification devolves to
a simple binary decision, rather than the need for a classifier
with multiple categories. This approach has the benefit of being
inclusive in that no motion mode is left out.
Examples of typical symmetrical motion modes include:
•
Texting: The mobile device is held in front of the user
while walking;
•
Phoning: The mobile device is held close to the head
while walking;
•
Heavy handbag: The mobile device is put in a heavy
bag (so the hand of the pedestrian is not swinging)
while walking;
•
Shirt pocket: The mobile device is put in a shirt pocket
while walking;
•
Static: The mobile device is not moving, regardless of
placement.
Similarly, some typical asymmetrical motion modes include:
•
Hand Swinging: The mobile device is held in a
swinging hand while walking;
2014 International Conference on Indoor Positioning and Indoor Navigation, 27th -30th October 2014
Phoning
Swinging
Trousers pocket
Back pocket
Orientation
Vertical
Pitch
Acceleration
Texting
Symmetrical
Fig. 4.
Asymmetrical
Acceleration and orientation signals of some typical symmetric and asymmetric motion modes.
•
Trouser pocket: The mobile device is put in a trouser
pocket (front or back) while walking;
•
Belt mounted: The mobile device is fastened to the
belt while walking;
•
Handshaking: The mobile device is (possibly periodically) shaken or tapped while not moving.
To determine whether a device placement is leading to
symmetric or asymmetric coupling, we examine both the
vertical acceleration and the pitch (rotation around the y axis)
component of the device orientation. This is the main reason
that long term tracking of the orientation of the device is so
important and why the RIOT technique introduced in Sec. III
is crucial to accurate classification. Based on extensive experimental evidence, we make the novel observation that symmetric motion modes lead to strong periodicity in the vertical
acceleration, closely approximating a sinusoidal wave, whereas
the pitch waveform is aperiodic. On the other hand, asymmetric
motion leads to aperiodicity in vertical acceleration, but strong
periodicity in pitch angle. This is shown graphically in Fig. 4,
illustrating a few examples of symmetrical and asymmetrical
motion modes. Notice that all symmetrical motion modes have
similar periodic vertical acceleration signals, with a period
corresponding to a single step, while all asymmetrical motion
modes have similar periodic pitch patterns, with a period
corresponding to a stride.
Thus, our classifier is based on assessing which one of
the vertical acceleration or pitch waveforms has the strongest
periodic component. More precisely, we compute the DFT of
each signal (pitch and acceleration) over the frequency range
that corresponds to human walking, typically between 1.2 Hz
and 2.5 Hz. We then use an energy detector to decide whether
the device placement is resulting in symmetric (vertical acceleration has stronger periodicity than pitch) or asymmetric
(pitch has stronger periodicity than vertical acceleration).
Based on this decision, the waveform with stronger periodicity is fed as input to an enhanced zero crossing detector that
has hysteresis to reject low amplitude noise. Note that if the
pitch waveform is used as input, each period of the waveform
actually corresponds to a stride of the user. Each step thus
is a half cycle of the periodic waveform. The candidate steps
are then passed to the next block, REPUT, described in detail
below.
B. REpetitive Pattern Update (REPUT)
Although the rotationally invariant orientation tracking algorithm described in Sec. III can achieve long term orientation
tracking of the device, it is not sufficient to accurately estimate
acceleration on the horizontal plane, which is important for
step validation. The major reason is that a very small tilt of
the device which cannot be corrected by the robust orientation
tracking with the noisy measurements from low-cost IMU
sensors, can significantly impact the acceleration signal on
the horizontal plane. For instance, a tilt error of only two
degrees changes the acceleration reading in the vertical axis
from 9.806 m/s2 to 9.801 m/s2 , but it can offset the measured
acceleration on the horizontal plane (either X or Y) by 0.342
m/s2 . To determine whether a step is valid or not (described in
detail below in Sec. IV-C), the raw acceleration measurements
are integrated twice to calculate the displacement of the step.
Any orientation offset is significantly amplified through double
integration which results in significant errors in displacement
estimation. The example two degree tilt error described above
can result in a cumulative displacement error of up to 20 metres
within 10 seconds.
To address this problem, we have further explored the physical constraints imposed by human walking. It is demonstrated
in [19] that humans have surprisingly consistent walking
patterns for consecutive steps. Inspired by this phenomenon,
we present an algorithm to estimate and correct the orientation
error using the repetitiveness in walking patterns, called repetitive pattern update (REPUT). Specifically, we assume that the
extrinsic states of the mobile device (velocity, height, etc.)
are largely the same after each step (or every stride in the
case of asymmetric motion). If we assume that the bias in the
orientation drifts slowly and remains constant over the duration
of the step, our task is to find the inverse rotation matrix R that
will compensate the offset, subject to the imposed constraints.
Consider the case of maintaining the velocity of the device
as a state that remains constant over a step. Suppose one
step starts at time t0 with velocity v 0 and ends at time te
with velocity v e . If there is no orientation bias, the difference
between the initial and terminal velocities should be zero i.e.
∆v = v e − v 0 = 0.
We can formulate the relation between the initial and
2014 International Conference on Indoor Positioning and Indoor Navigation, 27th -30th October 2014
Acc (<50%)
Acc (>50%)
Heading
Before REPUT1
2
After REPUT
0.5
1
1
0
0
0
−0.5
−1
−1
−1
8.5
−1.5
9
9.5
X: Time (s)
9
9.5
Y: Time (s)
−2
2
0
−2
0.5
0
−0.5
−2
8.5
0
Mean (Before) 2 Mean (After)
Y (m/s2)
Acceleration (m/s2)
2
8.5
9
9.5
Z: Time (s)
−2
−1
0
X (m/s2)
1
(a) The acceleration signals before and after REPUT
(b) Heading estimation
Fig. 5. Examples demonstrating (a) how the mean acceleration has drifted significantly from zero without the aid of REPUT and corrected to have mean value
close to zeros after REPUT has been applied, and (b) how the heading estimation calculates the best fit line for the instantaneous acceleration in the horizontal
plane.
terminal velocities as
ve = v0 +
te
X
R · at · ∆t,
(2)
t=t0
in which R is the inverse rotation matrix that corrects the
orientation drift, at = [axt , ayt , azt ] is the acceleration vector at
time t in earth’s coordinate frame, and ∆t is the time interval
between two consecutive acceleration readings. Then based on
our REPUT assumption, we have
∆v = v e − v 0 = R
te
X
·at · ∆t = 0,
(3)
t=t0
which allows us to solve for the unknown rotation matrix.
Other repetitive patterns which impose constraints on orientation drift can also be derived. For instance, if it is known
that the device is only used in 2D positioning, then the mobile
device has the same height above the ground at the beginning
and end of a step. Obviously in 3D tracking this cannot be used
without knowledge of the map as a user could be climbing
stairs for example, leading to a height difference. This can be
formulated as
!
te
t
X
X
z
at ∆τ ∆t = 0.
(4)
∆l = R
t=t0
where
azt
τ =t0
is the vertical acceleration along the Z axis.
These multiple repetitive constraints can be used to find
the optimal rotation matrix through least squares optimization.
To test the effectiveness of the REPUT algorithm, we have
conducted a simple experiment during which the pedestrian
walks normally with device held in a hand. Fig. 5(a) shows
the acceleration signals before and after applying the REPUT
algorithm. We can observe that the acceleration signals along
the X and Y axis are significantly biased without the REPUT
algorithm, whereas the offset has been corrected with the
REPUT algorithm.
C. Step Validation
The last block of the step detector algorithm is to determine
whether a candidate step really corresponds to a true step or
whether it could have been caused by some other motion which
resembles the waveform of a step. Essentially, this block acts
to remove false positives i.e. when a step has been detected but
it is not a real step of the user. To determine whether a step is
real or not, we again turn to basics of human motion. Simply,
a true step by the user must result in a forward displacement
within the range that corresponds to human step length. To
determine the displacement, it is necessary to double integrate
the raw acceleration measurements. This motivates the need for
extremely accurate measurements of orientation, provided by
the RIOT algorithm, and removal of orientation bias, provided
by the REPUT algorithm.
The cross correlation between the acceleration signal and
the orientation signal is also identified as an effective feature to
reject false positives because the acceleration and orientation
signal tend to have a higher correlation while shaking the
device than normal walking behaviour.
Then the features including horizontal/vertical displacements and cross correlation value are then feed into a simple
three nodes decision tree to determine whether the candidate
step really corresponds to a true step.
V.
ACCURATE HEADING ESTIMATION
In most existing work, the heading of the user is assumed
to be parallel to one axis of the device because the device
placement is assumed to be known or can be inferred from
sensor data. However, even with a simple placement like
texting, the heading of the device is not necessarily the same as
the heading of the user and differs greatly for different users.
With unconstrained placement of the device, the heading
of the device can be very difficult to determine, and can
dynamically change over time. However, since we have very
accurate orientation measurements of the device through RIOT
(see Sec. III) and REPUT (see Sec. IV-B), we can infer the
heading of the user by determining the principal swing axis,
which is aligned with the forward motion of the user. If we
consider the horizontal components of the acceleration over a
single step and project them into a two dimensional plane, the
best fit line (using simple least squares regression) corresponds
to the overall direction of forward motion. This leads to a
180◦ ambiguity where the sign of the heading angle is not
known. This is easily addressed by considering the trajectory
of the first half of the acceleration signals, corresponding to
the forward swing. Typical examples of the heading estimation
are shown in Fig. 5(b). Note that if the residual error is
small (Fig. 5(b) top and middle) we use the resulting estimate
as the offset between device and user heading; otherwise
(Fig. 5(b) bottom), we use the most recent offset that was
reliably estimated.
2014 International Conference on Indoor Positioning and Indoor Navigation, 27th -30th October 2014
VI.
10
E VALUATION
In this section, we compare and contrast the performance of
our approach with leading PDR approaches for step detection
and counting.
8
Error (%)
The competing approaches we compare against are:
Peak detection (PD:) The toe-off and heel-strike events
are usually associated with sharp changes in vertical or the
magnitude of acceleration which are the targets for typical peak
detection algorithms [4], [5], [11]. However, each heel-strike
could come with multiple peaks, especially when the mobile
device is in a swinging hand or attached low, e.g. a trousers
pocket, which significantly increases the algorithm complexity.
We use a low pass filter followed by a peak detector.
Zero-crossing (ZC): Another way of identifying the cyclic
patterns in acceleration is the detection of zero crossings [2],
[6], [20]. This algorithm is simple to implement, but can suffer
from false positive zeros crossings from events other than
walking.
Spectral analysis (STFT and CWT): This promising technique has attracted a large amount of attention recently due to
its robustness [3], [15], [17], [18]. Algorithms of this category
first convert the acceleration signal to frequency domain using
different algorithms like Fourier transform [15], [17], short
time Fourier transform (STFT) [3], or continuous wavelet
transform (CWT) [18], and then the dominant peak of the
signal in frequency domain is identified as the step frequency.
Windowed acceleration signal (magnitude) is common choice
for people working with spectral analysis [15]. As a result, online step detection makes this algorithm very computationally
intensive.
Auto/cross correlation (AC): Since each individual human
has a surprisingly consistent walking pace [19], the strong
periodicity of the acceleration signal from such walking behavior makes it possible to extract the cyclic pattern with
mean-adjusted autocorrelation of the sensor data sequence
or cross correlation with prepared templates – such as a
sample sequence of sensor data from the training phase [26].
ZEE, a recent approach that we use as comparison, can
also exploit characteristic waveforms from different device
placements [16].
Hand
Glass
Watch
Shirt
Trousers
Backpack
6
4
2
0
ZC
PD
STFT CWT
Algorithms
AC
R−PDR
Fig. 6. Comparison between state-of-the-art step detection algorithms and the
proposed approach, R-PDR, in terms of step detection accuracy for various
device positions.
from an experiment where a mix of daily behaviours such
as walking, standing, sitting, sending a text, making a call
and typing were performed. This is in comparison to other
work where a limited subset of behaviours (e.g. walking and
standing still) are used to demonstrate step detection accuracy.
The experiment lasted two hours and a total of 4669 true steps
were counted by the two foot mounted IMUs. Fig. 6 shows
the results for different algorithms and different placements. It
is interesting to note that for the competing algorithms, they
typically attain good performance for some placements, but
for one or two placements the accuracy is particularly poor,
in some cases as high as 10%. In particular, the handheld
placement typically registers the highest errors as repetitive
motion such as typing has strong periodicity like a true step.
Out of the competing algorithms, the autocorrelation (AC) algorithm performs reasonably well across the placements, with
a maximum error of 3%. The strength of our approach, R-PDR,
is apparent from the results, as regardless of placement, errors
are consistently below 1%. This experiment using realistic
human behaviour demonstrates how existing techniques suffer
from sensitivity to device placement, whereas our approach is
able to correctly detect the correct number of steps.
A. Experimental setup
For all tests, ground truth of steps was obtained using
foot mounted IMU’s (X-IO Technologies Inc.) which are
accurately calibrated. Two IMU’s were used, one on each
foot, to accurately count the number of steps. Ground truth of
position was obtained by placing numbered labels along the
route at 1-m evenly spaced intervals and using a down facing
hand-held camera to note the exact time when the foot crossed
the midpoint of the label. For each test, the user was outfitted
with 2 IMU’s (one on each foot) and 6 Nexus-S smartphones
capturing IMU data at various positions of the body. This
allows us to perform direct comparisons of accuracy, as an
accurate PDR system should report the same number of steps,
regardless of device placement.
B. Overall detection accuracy
In this test, we examine the accuracy of overall step
detection for various placements for each of the competing
algorithms, compared to the ground truth. Data was obtained
C. False positive rejection
In this test, using data from a number of behaviours that
are repetitive but do not correspond to forward displacement,
we demonstrate the performance of existing algorithms in
determining whether a step has actually been taken or not.
This is shown in Fig. 7. Each behaviour, such as tapping, was
performed 200 times.
Existing algorithms which are based solely on characteristics of the waveform itself, such as periodicity, are unable
to distinguish between behaviours like tapping and true steps
caused by walking. This is because they are only designed
to detect periodical patterns and do not consider whether a
motion, as registered by the sensors, actually corresponds to
true forward motion. This is made even more challenging
for unconstrained device placement. The suite of techniques
presented in R-PDR allows us to evaluate, for each step, the
displacement of the user and assess whether or not a step has
really been taken. This ability to distinguish between true and
2014 International Conference on Indoor Positioning and Indoor Navigation, 27th -30th October 2014
250
Ground truth
Nodding
Foot tapping
R−PDR
PHA
30
4
200
20
2
150
100
10
Y (m)
Heading (rad)
Number of steps detected
Tapping
0
0
−10
−2
−20
50
−4
0
0
ZC
PD
STFT CWT
Algorithms
AC
50
100
Time (s)
−30
0
20
X (m)
40
R−PDR
Fig. 7. Comparison between state-of-the-art step detection algorithms and the
proposed approach, R-PDR, in terms of step detection accuracy for various
device positions.
false steps makes our algorithm greatly outperform existing
algorithms.
D. Heading estimation
Existing heading algorithms assume that the heading of the
user is parallel with a sensor axis e.g. that the phone is carried
in a hand facing forward, called parallel heading assumption
(PHA). In this test, we demonstrate how this assumption can
lead to large heading errors if the device is rotated even a
small amount. The user carried the device facing forward.
Halfway through the test, the device was rotated in the hand by
30◦ . The results are shown in Fig. 8. This demonstrates that
the trajectory estimated with the existing technique diverges
significantly from the ground truth when we slightly change
the orientation of the device whilst orientation changes do
not impact the proposed approach. This ability to accurately
estimate heading, regardless of device orientation, is extremely
important for long term positioning, as small errors in heading
accumulate rapidly and lead to the failure of PDR. Note that
the initial heading is estimated from magnetometer and thus
biased due to the magnetic distortions in indoor environments.
E. Impact on positioning systems
Lastly, we examine how errors in the PDR couple through
to the positioning system. We use MapCraft, a state of the art
technique for converting noisy dead reckoned trajectories into
accurate locations. MapCraft does not require an initial starting
point or heading and tolerates large errors in dead reckoned
data. Using this technique as a common benchmark, we assess
how different PDR step and heading estimation techniques
impact overall positioning error.
This experiments were conducted in an office building
shown in Fig. 10. During the experiments, the subjects were
mounted with several devices simultaneously in different parts
of the body, typically hand, watch, glasses, and trousers pocket.
Then they walk anywhere in the building without planned
routes, to realistically capture real pedestrian motion, rather
than artificial, constant speed trajectories. They were told to
move freely and may have different motion modes including
walking, standing still, tapping, foot tapping, typing, sitting
down, bending to pick up something on the floor, etc. To
provide accurate ground truth, numbered labels were placed
Fig. 8.
Comparison of different heading estimation techniques.
along corridors and within rooms on a 2 m grid. Using the
device’s camera, these were filmed at the same time experiments were conducted. The time-synchronized video streams
were then mapped to locations on the floorplan.
Fig. 9 compares the step detection accuracy of R-PDR
against autocorrelation which is reported to have the best
overall detection accuracy in existing techniques. It is observed
that R-PDR always outperforms autocorrelation for various
human motions and placements of devices. The major reason
lies in the fact that R-PDR could reject false positives while
AC cannot. Motions like nodding, tapping, typing, and foot
tapping can generate step-like signals without really moving
forward. As a result, AC significantly over counts the number
steps, which, inevitably leads to significant errors in tracking
(the significant drift of trajectory in Fig. 10(e)), even with the
state-of-the-art map matching algorithm (Fig. 10(c)). Fig. 10(f)
shows the error distribution function of the R-PDR and the
competing approaches. It is observed that R-PDR has a RMS
error of 0.98m regardless of device placement. An example of
the testing trajectory can be found in Fig. 10.
VII.
R ELATED W ORK
A wide variety of algorithms have been developed to identify steps from inertial signals so far [8]. Typical step detection
algorithms have been developed to accurately identify specific
events for data segmentation, for instance, peaks, zero crossing,
etc. These algorithms are mostly applied to acceleration or
angular velocity signal, or their combinations [18], [22]. To
handle unconstrained device placement, a number of techniques have been proposed, such as motion recognition which
attempts to classify different motion patterns (such as hand
swinging, texting, running, etc.) based on waveform features
[23], [25]. Rather than building a rich classifier, which requires
significant training, we take a simpler approach and consider
a novel observation based on the periodicity of orientation and
acceleration signals. To the best of our knowledge, this is the
first technique that considers both acceleration and orientation
waveforms simultaneously.
The most crucial underlying assumption that existing algorithms make is that people do not generate periodical
acceleration signals while not walking, such as a hand swing
without the user really moving forward. This is because the
two actions (a step and a hand swing) can actually generate
2014 International Conference on Indoor Positioning and Indoor Navigation, 27th -30th October 2014
0
5 10
0
m
(a) Ground truth
(b) R-PDR + map matching
1
0.8
−20
−20
0.6
−30
−40
0.4
−60
−40
Fig. 10.
CDF
0
X (m)
−10
−50
m
(c) AC + PHA + map matching
0
X (m)
5 10
0.2
−80
−20
−10
0
10
Y (m)
20
30
R−PDR + map matching
AC + PHA + map matching
MapCraft (handheld)
−80
−60
−40 −20
Y (m)
0
20
0
0
(d) R-PDR
(e) AC + PHA
Experiments in the office environment, showing the ground truth and reconstructed trajectories.
an identical acceleration signal patterns, especially when the
step detection algorithms are applied to the magnitude of
the acceleration [1]. The situation would be exacerbated with
less constrained motions like crawling, stumbles, side-steps,
shuffles, etc. [8], which pose great challenge to motion recognition algorithms. We demonstrate that by accurately estimating
the horizontal and vertical displacement of each step (which
requires enhanced filtering to prevent drift), we can accurately
determine whether a forward step has been taken.
Orientation tracking/heading estimation, tackles the problem of continuously estimating the orientation of the mobile device, relative to North. Typical orientation tracking
algorithms with inertial data are based on indirect Kalman
filters [21] or variants of Kalman filters like extended Kalman
filters (EKF) [27] or Unscented Kalman filters (UKF) [10]. The
standard way of tracking the orientation with Kalman filters
is to update the orientation using the angular velocity from
gyroscope and compensate for the long term drift of orientation
tracking from noisy gyro data with the magnetic field from
magnetometer or gravity vector from accelerometer.
However, the majority of research on orientation tracking
in the context of indoor positioning is to determine the heading
of the pedestrian rather than the heading the device. Most
research therefor assumes the position where the mobile device
is attached is known, thereby making the assumption that
the heading of the pedestrian is always consistent with the
heading of the mobile device [12]. This is not always true.
Recent research investigating the difference between the user’s
heading and the phone’s heading [16] demonstrated that the
second harmonic of the acceleration is either completely absent
or is extremely weak in the direction perpendicular to the user’s
walk. It is conversely dominant in the direction parallel to
the user’s walk. Moreover, the long term drift of orientation
tracking is still a pressing problem. Though in existing work
10
20
Error (m)
30
(f) Comparison of tracking errors
the magnetic field of the earth can mostly compensate long
term yaw errors, it has little correction power in roll and pitch
estimation. This is because the gravity vector can be used only
when the device is static [7].
VIII.
C ONCLUSIONS
Pedestrian dead reckoning is a technique that can act as
the foundation of an accurate indoor positioning system by
exploiting low cost, infrastructure free device sensors. Work
to date has made great strides in demonstrating the potential
of PDR, but subject to constraints on device placement.
These constraints are limiting the widespread adoption of
PDR. Rather than seeking a placement specific approach, or
attempting to derive the placement of the device, we have
taken an alternative approach in this work and presented a
suite of techniques that together provide robust PDR (RPDR). Through extensive experiments using R-PDR, we have
demonstrated that step detection and heading estimation are
accurately performed even when the position of the device is
changed dynamically through the test. In addition, we also
showed how R-PDR is able to reject common behaviours like
typing that appear like steps to competing PDR techniques.
We believe that the techniques presented here will have wide
applicability not only in consumer devices like smartphones
but also in wearable devices like smartwatches and glasses.
In summary, R-PDR is a comprehensive technique that will
enable the adoption of accurate, reliable and robust indoor
positioning, without requiring any calibration or tuning.
Acknowledgements The authors would like to thank the
anonymous reviewers for their comments and suggestions to
improve the paper. They also acknowledge the support of the
EPSRC through grants EP/L00416X/1.
2014 International Conference on Indoor Positioning and Indoor Navigation, 27th -30th October 2014
Hand
[7]
[8]
[9]
[10]
Glasses
[11]
[12]
[13]
Watch
[14]
[15]
[16]
Trousers
[17]
[18]
[19]
[20]
[21]
Fig. 9. Typical human motions and the corresponding step detection accuracy
for R-PDR and AC.
[22]
R EFERENCES
[1]
M. Alzantot and M. Youssef. UPTIME: Ubiquitous pedestrian tracking
using mobile phones. In Proc. IEEE Wirel. Commun. Netw. Conf.
(WCNC’12), pages 3204–3209. Ieee, Apr. 2012.
[2]
S. Beauregard and H. Haas. Pedestrian dead reckoning: A basis for
personal positioning. In Proc. 3rd Workshop Pos. Nav. Commun.
(WPNC’06), pages 27–36, 2006.
[3]
A. Brajdic and R. Harle. Walk detection and step counting on unconstrained smartphones. In Proc. ACM Conf. Ubi. Comput. (UbiComp’13),
pages 225–234, Zurich, Switzerland, 2013. ACM Press.
[4]
[5]
[6]
I. Bylemans, M. Weyn, and M. Klepal. Mobile Phone-Based Displacement Estimation for Opportunistic Localisation Systems. In Proc. 3rd
Int. Conf. Mob. Ubi. Comput. Syst. Services Technol. (UbiComm’09),
pages 113–118, Sliema, Oct. 2009.
L. Fang, P. Antsaklis, L. Montestruque, M. B. McMickell, M. Lemmon,
Y. Sun, and H. Fang. Design of a wireless assisted pedestrian dead
reckoning system-the NavMote experience. IEEE Trans. Instrument.
Measurement, 54(6):2342–2358, 2005.
P. Goyal, V. J. Ribeiro, H. Saran, and A. Kumar. Strap-down Pedestrian
[23]
[24]
[25]
[26]
[27]
Dead-Reckoning system. In Proc. Int. Conf. Indoor Pos. Indoor Nav.
(IPIN’11), pages 1–7. Ieee, Sept. 2011.
T. Harada, T. Mori, and T. Sato. Development of a Tiny Orientation
Estimation Device to Operate under Motion and Magnetic Disturbance.
The Int. J. Robotics Res., 26(6):547–559, June 2007.
R. Harle. A Survey of Indoor Inertial Positioning Systems for Pedestrians. IEEE Commun. Surveys & Tutorials, 15(3):1281–1293, Jan. 2013.
S. Hemminki, P. Nurmi, and S. Tarkoma. Accelerometer-based transportation mode detection on smartphones. In Proc. 11th ACM Conf.
Embedded Netw. Sensor Syst. (Sensys’13), pages 1–14, New York, NY,
USA, 2013.
B. Huyghe, J. Doutreloigne, and J. Vanfleteren. 3D orientation tracking
based on unscented Kalman filtering of accelerometer and magnetometer data. In Proc. IEEE Sens. App. Symp. (SAS’09), pages 1–5, New
Orleans, LA, USA, 2009.
J. W. Kim, H. J. Jang, D. H. Hwang, and C. Park. A step, stride and
heading determination for the pedestrian navigation system. Journal of
Global Position. Syst., 3(1):273–279, 2004.
F. Li, C. Zhao, G. Ding, J. Gong, C. Liu, and F. Zhao. A reliable and
accurate indoor localization method using phone inertial sensors. In
Proc. ACM Conf. Ubiquitous Comput. (UbiComp’12), pages 421–430,
2012.
H. Lu, J. Yang, Z. Liu, N. D. Lane, T. Choudhury, and A. T. Campbell.
The Jigsaw continuous sensing engine for mobile phone applications. In
Proc. 8th ACM Conf. Embedded Netw. Sensor Syst. (SenSys’10), pages
71–84, New York, New York, USA, 2010. ACM Press.
D. Mizell. Using gravity to estimate accelerometer orientation. In Proc.
7th IEEE Int. Symp. Wearable Computers (ISWC’03), pages 252–253.
Ieee, 2003.
J.-g. Park, A. Patel, D. Curtis, S. Teller, and J. Ledlie. Online pose
classification and walking speed estimation using handheld devices. In
Proc. ACM Conf. Ubi. Comput. (UbiComp’12), pages 1–10, New York,
New York, USA, 2012. ACM Press.
A. Rai and K. Chintalapudi. Zee: Zero-effort crowdsourcing for
indoor localization. In Proc. 18th Ann. Int. Conf. Mob. Comput. Netw.
(MobiCom’12), pages 1–12, Istanbul, Turkey, 2012.
N. Ravi, N. Dandekar, P. Mysore, and M. Littman. Activity recognition
from accelerometer data. In Proc. 7th Conf. Innov. App. Artificial Intell.
(AAAI’05), pages 1541–1546, 2005.
V. Renaudin, M. Susi, and G. Lachapelle. Step length estimation using
handheld inertial sensors. Sensors, 12(7):8507–8525, Jan. 2012.
J. Rose and J. G. Gamble. Human Walking. Lippincott, Williams and
Wilkins, Baltimore, PA, USA, 3rd edition, 2006.
S. H. Shin, M. S. Lee, and P. C. G. Pedestrian dead reckoning system
with phone location awareness algorithm. In Proc. IEEE/ION Position
Location Nav. Symp. (PLANS’10), pages 97–101, 2010.
Y. S. Suh. Orientation Estimation Using a Quaternion-Based Indirect
Kalman Filter With Adaptive Estimation of External Acceleration. IEEE
Trans. Instrument. Measurement, 59(12):3296–3305, Dec. 2010.
M. Susi, V. Renaudin, and G. Lachapelle. Motion mode recognition and
step detection algorithms for mobile phone users. Sensors, 13(2):1539–
62, Jan. 2013.
J. S. Wang, C. W. Lin, Y. T. Yang, and Y. J. Ho. Walking pattern
classification and walking distance estimation algorithms using gait
phase information. IEEE Trans. Biomedical Engineer., 59(10):2884–
2892, 2012.
Z. Xiao, H. Wen, A. Markham, and N. Trigoni. Lightweight map
matching for indoor localization using conditional random fields. In
Proc. Int. Conf. Info. Process. Sensor Netw. (IPSN’14), Berlin, Germany,
2014.
J. Yang. Toward physical activity diary: motion recognition using simple
acceleration features with mobile phones. In Proc.1st Int. workshop
Interactive multimedia for consumer electronics, pages 1–9, Beijing,
China, 2009.
H. Ying, C. Silex, and A. Schnitzer. Automatic step detection in
the accelerometer signal. In Proc. 4th Int. Workshop Wearable and
Implantable Body Sensor Netw. (BSN’07), pages 80–85, 2007.
X. Yun and E. R. Bachmann. Design, Implementation, and Experimental
Results of a Quaternion-Based Kalman Filter for Human Body Motion
Tracking. IEEE Trans. Robotics, 22(6):1216–1227, Dec. 2006.