=Paper=
{{Paper
|id=Vol-2498/short50
|storemode=property
|title=Daily action dead reckoning using smartphone sensors
|pdfUrl=https://ceur-ws.org/Vol-2498/short50.pdf
|volume=Vol-2498
|authors=Takumi Otsuki,Kohei Kanagu,Kota Tsubouchi,Nobuhiko Nishio
|dblpUrl=https://dblp.org/rec/conf/ipin/OtsukiKTN19
}}
==Daily action dead reckoning using smartphone sensors==
https://ceur-ws.org/Vol-2498/short50.pdf
Daily Action Dead Reckoning Using Smartphone
Sensors
Takumi Otsuki1 , Kohei Kanagu2 , Kota Tsubouchi3 , and Nobuhiko Nishio4
1
Ritsumeikan University, Shiga, Japan
tsukky@ubi.cs.ritsumei.ac.jp
2
Yahoo Japan Corporation, Tokyo, Japan
kingu@ubi.cs.ritsumei.ac.jp
3
Yahoo Japan Corporation, Tokyo, Japan
ktsubouc@yahoo-corp.jp
4
Ritsumeikan University, Shiga, Japan
nishio@cs.ritsumei.ac.jp
Abstract. Pedestrian dead reckoning (PDR) is easy to introduce be-
cause it requires no equipment for the environment. PDR results can
provide an atomic physical behavior such as step detection and turn-
ing in walking, however providing a lexible response to a user’s daily
actions other than walking like sitting, moving in a line or standing is
tough. Our research objective is to make PDR more usable in spite of
these daily actions by estimating the movement situation using the same
sensor information as conventional PDR. We applied a state transition in
the movement situation recognition using the transition restrictions ex-
isting between the moving situations. A new method for moving context
recognition using machine learning with accelerometer and gyroscope fea-
ture vectors simultaneous with PDR is proposed herein. The experiments
show that, the movement situation recognition and the state transition
technique are eicient for the whole dead reckoning performance. The
proposed method can also be applied to improve the PDR positioning
accuracy.
1 Introduction
Positioning techniques have been actively studied in the recent years with the
spread of services using location information. Indoor location information is use-
ful because it can be used for working management of factory workers and nav-
igation in a complex urban indoor environment, among others. Pedestrian dead
reckoning (PDR) is a typical positioning method that is functional in both indoor
and outdoor environments and achieved by continuously estimating the walk-
ing conditions, distance and direction using the accelerometers, gyroscope and
magnetometer of the user’s smartphone, and assimilating these measurements to
estimate the current position. Moreover it does not depend on the environment
thereafter and does not require extra equipment.However, the problem is that
positioning errors are accumulated. Since PDR is a relative positioning method,
2 Otsuki et al.
no absolute location calibration arrives. Therefore, our research focuses on the
user’s most of all kinds of movement not only walking steps that may cause
any positioning error. The current PDR, only focuses on walking recognition the
other body action states of the user are hardly considered. An erroneous step
detection when sitting in a chair at the same place, which is a daily action often
seen indoors is conirmed because of the changing leg or body posture and one’s
orientation even though he/she does not move at all. For example, we frequently
perform a changing movement in posture and body orientation, stay at the same
place with some steps , or sit for a long time as well as perform irregular walk-
ing in which stopping is frequently and forcibly repeated such as in the line or
during congestion (hereinafter referred to as non-moving state). These states
include motion without movement, which leads to a false detection of walking
and direction estimation in PDR. The sitting condition requires to distinguish it
from the standing condition which may start moving at any time. The machine
learning technique is efective in recognizing such irregular actions and states. In
order to improve the positioning accuracy, some methods perform movement or
state identiication using highly expensive and high-precision sensors such as the
Inertial Measurement Unit(IMU)[3]; however, such devices are not available for
smartphones. The other methods using extra sensors and ixing them at speciic
positions for all would bring many limitations. Aside from clarifying the user’s
moving state, this study also performs PDR positioning by simultaneously per-
forming state/motion recognition using the learned model and PDR positioning
and algorithm optimization according to the estimated result of the model. We
propose a new positioning method called daily action dead reckoning to minimize
the accumulated errors and improve the positioning accuracy.
2 Related work
2.1 Activity recognition using a smartphone
Makita et al.[3] determined 11 types of walking motions and states including
a context that does not follow the movement that is diicult to predict based
on only tracked walking trajectory(e.g.sitting motions) using a sensor device
attached to the waist.Their recognition achived with 90% accuracy or more.
However, it is burdensome for the user to wear the sensor on a part of the body.
As a method without a sensor device attached to the body, Pei et al.[4] used
a smartphone in a pocket to recognize bending considering the walking speed,
walking state, running state, and stop. Whether Cosukun et al.[5] proposed a
method that recognizes the walking(e.g. climbing stairs) and running conditions.
KHAN et al.[6] combined 15 smartphone acceleration sensors, microphones and
barometric pressure sensors to identify 15 daily activities and conditions. All
of the previous studies identiied the way of walking, etc., which was diferent
from the user’s movement state recognition that is the purpose of the present
research.
Daily Action Dead Reckoning Using Smartphone Sensors 3
2.2 PDR using another sensor
One method uses the IMU as a method of utilizing other sensors. The IMU is
a very high accuracy sensor compared to the micro electro mechanical system
mounted on a general smartphone and often used for PDR. Coskun et al.[5] used
an IMU attached to the foot to estimate the leg angle etc. and optimize the PDR
according to the angle to reduce the positioning error of the PDR. Park et al.[2]
classiied the user’s movement and condition into eight using an IMU attached
to the leg and used it for PDR positioning. However, the IMU is a very expensive
sensor, and it has many limitations such as the need to ix the holding state to
a part of the body.
2.3 PDR using the walking context recognition results
Qian et al.[7] estimated the attitude of the smartphone held by a user and iden-
tiied the holding state of the smartphone such as calling using the smartphone,
holding it putting it in the pocket and typing characters. Depending on the
identiication result, they then optimized the PDR algorithm and improved the
positioning accuracy. However,this method cannot dynamically change the al-
gorithm and works only with a single algorithm. Tian et al.[8]also estimated the
attitude of the smartphone held by a user including holding the smartphone,
swinging, and putting it in the pocket. They subsequently optimized the PDR
algorithm according to the identiication result. Kanagu et al.[9]classiied the
state of searching for products generated during shopping from the appearance
of walking into three states, and aimed to improve the positioning accuracy.
Their method was limited because the target environment was super.
3 Daily action dead reckoning(DDR)
Fig. 1 shows the system low of the proposed method which was divided into the
learning and evaluation phases.
3.1 Learning phase
In the proposed method, we created a model by logistic regression. Logistic
regression is a regression model classiied as supervised machine learning having
the features that enable inding useful features because the weight of each feature
that leads to classiication is known. Normal walking is deined as ”walking”,
walking which is similar to stopping is deined as ”moored walk”(e.g, crowded).
Standing is deined as ”stop”, sitting is deined as ”sitting”. In addition, we also
identify standing and sitting movements for using the state transitions(3.1)
Evaluation phase In the evaluation phase, the user’s walking context is deter-
mined by analyzing the sensor data using a trained context-aware model. The
adopted t value is the same as that used in the learning phase. The extracted
4 Otsuki et al.
Fig. 1. Processing low of the proposed method
feature vectors are plotted on the data space, and a trained context recognition
model detects the context of the feature vectors. The identiied walking context
is merged with the ordinal PDR result to obtain the PDR result with context
information. The PDR algorithm is adapted to the PDR method proposed by
Yoshimi et al. [1] who accumulated changes in the stride and direction estimated
by the sensor data to identify the user’s indoor location at that time.
State transition Fig .2depicts the behavior and state identiied by the proposed
method and considered to be transitioning between some behaviors and states.
The sitting state is next to the sitting motion. The sitting motion is after the
sitting state. Some transitions cannot occur. For example, when the walking state
occurs after the sitting motion or the sitting state occurs after the leaving motion,
either of the discrimination results is erroneous. Therefore, if a class cannot be
transitioned, the state transition is performed to avoid making a transition. The
igure shows a state transition diagram that omits impossible transitions. In
the proposed method, the decrease in the discrimination accuracy between the
sitting and stop states can be by provided by identifying the action that triggers
the sitting state such as the sitting and standing actions. However, problems also
occur in this case. In the proposed method, the sensor value is cut out in time
series (the length of the cut-out time is taken as the window width) to calculate
the feature amount, but the window is appropriate in the state(e.g. sitting action
and walking state) and the width is diferent. It is necessary to return the window
width for short continuous operation and long continuous condition. If the length
is short, suicient features for walking can not be acquired, and if the length is
long, the motion may be buried. In the proposed method, state transition is
done by creating two classiiers with diferent window widths and integrating
the classiication results.
Daily Action Dead Reckoning Using Smartphone Sensors 5
Fig. 2. State transition
Optimization of the PDR algorithm A state of moving straight ahead to-
ward a destination is assumed for walking. This was the method used by Yoshimi
et al[1]which was adopted as a normal walking algorithm. In moored walk, in
a crowd or in a line, walking and stopping are continuous, and a stride may
become short;therefore a iltering algorithm that ignores a certain amount of
change in the movement direction is omitted. The stop assumes that the vehicle
stops and does not move;thus it does not use Yoshimi et al’s gait detection algo-
rithm and does not change the user’s current location. Similarly, step detection
is turned of while sitting down. However while sitting down, the direction of
the body does not signiicantly change from the time when it irst sat;therefore
it maintained the direction when facing the same direction for a ixed time or
more. When away from the seat reset to the direction it was holding.
4 Evaluation
4.1 Experiment condition
In this evaluation, an attempt was made to recognize the state in which the
user is moving in the experimental building. Several chaise lounges were used
in the experimental environment. The subject was asked to walk, move in a
row, sit down, and stop. The subject in their twenties walked a speciic route
around the laboratory. The subject sat on a bench on the route, asked to change
pace at any time, and was allowed to move his posture, stride, foot, etc. As a
means of data collection, special sensor devices and sensors are not suitable for
practical use;thus we used the common smartphone Nexus 5 for evaluation. The
smartphone is put in the pocket of the pants in consideration of the burden on
the user.
6 Otsuki et al.
4.2 Estimation accuracy of models and optimization results of PDR
For the feature quantities of the test data, the weighting factors in each classiier
were added up. The result of the classiier with the largest weighting factor value
was adopted as the inal estimation result. The weighting factor for the feature
value changed with each learning;thus 10 classiiers were created for each label,
and the weighting factor values were summed. In order to compare the estimation
accuracy using multiple classiiers performed by the proposed method, the case
of using a single classiier is compared. The results of using a single classiier
and the F value of each class of the proposed method (with or without state
transition) are shown in the table, and the accuracy rate of each method is
shown.(1)
Table 1. f-Number of classiiers.
Context Conventional method Proposed method
Walk 0.45 0.80
Moored walk 0.33 0.76
Stop 0.36 0.70
Sit 0.21 0.80
Sit_down 0.64 0.72
Stand_up 0.74 071
When a single discriminator was used the window width was set to 250 ms,
and suicient features (steps 1 and 2) of walking and mooring walk cannot be
obtained. The identiication accuracy of stand and sit motions were high, but
the state identiication accuracy was low. However, in the proposed method that
created two classiiers, gait etc. could be captured by other classiiers. Moreover
the features could be captured, and the classiication rate of the state could be
improved. Therefore, the method using classiiers with diferent window widths
was efective, but the state transition did not improve the estimation accuracy
that much. This is considered to be caused by a mistake, which caused an er-
roneous transition by giving priority to the triggering of the leaving / sitting
motion. Therefore, it is thought that a mechanism such as a transition pending
state is needed so that it does not immediately transition when coming and go-
ing sitting behavior comes in the future, and decides the transition by looking
at some subsequent estimation results.The experiment for DDR was conducted
with the route and behavior as shown in the FIG3.
As a result, compared to the conventional method, the step detection error
and the direction estimation error at the time of stopping, sitting, and moored
walk are reduced(Fig4). During the stop, while the user is seated, the proposed
method does not move almost at all, and the moored walk has few positioning
errors due to a step estimation error. In the case shown in the igure, the efect
looks small because it is a short time experiment, but it can be said that it is
Daily Action Dead Reckoning Using Smartphone Sensors 7
Fig. 3. Walking motion for the evaluation experiment
necessary to optimize according to the user’s state like the proposed method,
assuming long time use.
Fig. 4. Tracked trajectory(left:proposed method, right:conventional method)
5 Conclusion
This study proposed a method that used sensor data such as acceleration and
angular acceleration that are commonly used in conventional PDR technology
to detect characteristic daily walking patterns and movements,judge the user’s
condition and use PDR. We proposed DDR to optimize the algorithm and reduce
the positioning error. Our future tasks aim to utilize this method in the actual
environment(e.g.oices) improve the estimation accuracy of the classiier, and
add motions and states to be identiied. As a state to be added, we think about
the movement to the side where the direction and movement direction of the
8 Otsuki et al.
body are diferent, and the movement to the rear. The PDR operates with the
same movement direction as the direction of the body, these movements cannot
be distinguished from normal walking, which causes positioning errors. We also
aim to discover new issues by operating the proposed method in an environment
closer to the actual usage environment.
Acknowledgement
References
1. Shun Yoshimi, Kohei Kanagu, Masahiro Mochizuki, Kazuya Murao, and Nobuhiko
Nishio. Pdr trajectory estimation using pedestrian-space constraints: Real world
evaluations. In Adjunct Proceedings of the 2015 ACM Interna- tional Joint Confer-
ence on Pervasive and Ubiquitous Computing and Proceed- ings of the 2015 ACM
International Symposium on Wearable Computers, Ubi- Comp/ISWC’15 Adjunct,
pp. 14991508, New York, NY, USA, 2015. ACM.
2. PARK, So Young; JU, Hojin; PARK, Chan Gook. Stance phase detection of mul-
tiple actions for military drill using foot-mounted IMU. sensors, 2016, 14: 16.
3. Koji Makita, Masakatsu Kourogi, Tomoya Ishikawa,Takashi Okuma, and Takeshi
Kurata. PDRplus: HumanBehaviour Sensing Method for Service Field Analysis,pp.
25f30. Springer Japan, Tokyo, 2014.
4. Ling Pei, Robert Guinness, Ruizhi Chen, Jingbin Liu,Heidi Kuusniemi, Yuwei
Chen, Liang Chen, and Jyrki Kaistinen. Human behavior cognition using smart-
phone sensors. Sensors, Vol. 13, No. 2, pp. 1402f1424, 2013.
5. D. Coskun, O. D. Incel, and A. Ozgovde. Phone posi-tion/placement detection
using accelerometer: Impact on activity recognition. In 2015 IEEE Tenth In- ter-
national Conference on Intelligent Sensors, Sensor Net- works and Information
Processing (ISSNIP), pp. 1f6, April 2015.
6. Khan, Adil Mehmood, et al. Activity recognition on smartphones via sensor-fusion
and kda-based svms. International Journal of Distributed Sensor Networks, 2014,
10.5: 503291.
7. QIAN, Jiuchao, et al. Continuous motion recognition for natural pedestrian dead
reckoning using smartphone sensors. In: Proceedings of the 27th International
Technical Meeting of the ION Satellite Division, ION GNSS. 2014.
8. TIAN, Qinglin, et al. An enhanced pedestrian dead reckoning approach for pedes-
trian tracking using smartphones. In: Intelligent Sensors, Sensor Networks and
Information Processing (ISSNIP), 2015 IEEE Tenth International Conference on.
IEEE, 2015. p. 1-6.
9. KANAGU, Kohei; TSUBOUCHI, Kota; NISHIO, Nobuhiko. Colorful PDR: Col-
orizing PDR with shopping context in walking. In: Indoor Positioning and Indoor
Navigation (IPIN), 2017 International Conference on. IEEE, 2017. p. 1-8.
10. S. O. H. Madgwick, A. J. L. Harrison, and R. Vaidyanathan. Estimation of imu and
marg orientation using a gradient descent algorithm. In 2011 IEEE Inter- national
Conference on Rehabilitation Robotics, pp. 1f7, June 2011.