=Paper=
{{Paper
|id=None
|storemode=property
|title=People Identification Based on Sitting Patterns
|pdfUrl=https://ceur-ws.org/Vol-960/paper7.pdf
|volume=Vol-960
}}
==People Identification Based on Sitting Patterns==
https://ceur-ws.org/Vol-960/paper7.pdf
People Identification Based on Sitting Patterns
Nguyen Gia 1 and Takuya Takimoto 2 and Nguyen Doan Minh Giang 3
and Jin Nakazawa 4 and Kazunori Takashio 5 and Hideyuki Tokuda 6
Abstract. This paper proposes a people identification method based it by simple actions. So, it is acceptable to inference to the user who
on the sitting patterns. This method uses weak evidences from pres- is sitting beforehand based on weaker evidences. Collecting weak
sure sensors, accelerometer sensors, and light sensors placed on a evidences also can be implemented without any psychological and
chair to recognize who is sitting on the chair without any psycholog- physical burden of users.
ical and physical burden on users. We discuss how we have imple- This paper proposes an easy deployment and inexpensive people
mented the system using softmax regression model, gradient descent identification method that uses weak evidences from pressure sen-
algorithm and nearest neighbor search algorithm. Our experimental sors, accelerometer sensors and light sensor placed on a chair. The
result shows that this method can be used in places which has private reason of using pressure sensor is the difference of weight among
properties such as a home or small a office. users. Also, we think that the sitting patterns are different between
users so we use accelerometer sensors to recognize the movement
of the chair when user sit in it. The light sensor is used to measure
1 Introduction the coverage of user in the chair. We have used softmax regression
model, a supervised learning algorithm and gradient descent algo-
Nowadays, there are several biometric people identification methods rithm, an algorithm to solve optimization problem to inference who
such as fingerprint based[1], iris based[2] or by the using of vein[3]. is sitting in the chair
These biometric identifiers are strong and suitable for applications Remainder of this paper is organized as follows. Section 2 de-
that request high accuracy such as security applications. However scribes the design and implementation of system. The softmax re-
these identifiers annoy user with the requests for specific actions. For gression model, gradient descent algorithm, nearest neighbor search
example, in the case of fingerprint, users have to properly touch a fin- algorithm and how they are used are discussed in Section 3. Section
gerprint scanner or in the case of retina, users have to look at retina 4 shows the result of our experiment while Section 5 is about related
scanner for a while, which might cause a psychological and physical work. Conclusions and future work are described in Section 6.
burden on users. These methods also need delicate and expensive de-
vices such as fingerprint scanner or retina scanner. In such situations
as inside of a house or a small office with a small number of users, we 2 Design and Implementation
do not need high accuracy as those available with strong identifiers. 2.1 Hardware
For example, in a small office, which employees come from some
different country, an employee comes to the office, sits on a public We use SunSPOT[8] for accelerometer sensor and light sensor.
chair and turns on a public computer. And then, a greeting sound of SunSPOT (Sun Small Programmable Object Technology) as shown
his/her country comes out of the speaker and that computer’s lan- in Figure 1(a) is a wireless sensor network mote which developed by
guage will be automatically change to his/her native language. Is it Sun Microsystems. One SunSPOT device has three types of sensor
interesting? For the other scenario, an office has a meeting but the including an accelerometer sensor, a light sensor and a temperature
boss is in a business trip so he/she uses a robot for teleconference. sensor. In this research, we only use one SunSPOT device to sense
The robot stands in the middle of meeting room and when the boss accelerometer and light data. These data can be sent to a computer
wants to talk with one of his/her employees, he/she only has to let for processing through a base station as shown in Figure 1(b).
the robot knows the employee’s name instead of rotating robot by
hand. Both of these scenarios can be realized with one of the above
people identification methods, but using biometric identifiers for this
scene is wasteful and unnecessary. Any mistake of people recogni-
tion in these scenes is not a big problem, users can easily overcome
1 FPT Software Co. Ltd, Vietnam, email: gia@ht.sfc.keio.ac.jp
2 Graduate School of Media and Governance, Keio University, Japan, email: (a) SunSPOT device (b) SunSPOT base station
tacky@ht.sfc.keio.ac.jp
3 Faculty of Environment and Information Studies, Keio University, Japan,
email: spider@ht.sfc.keio.ac.jp Figure 1. SunSPOT
4 Graduate School of Media and Governance, Keio University, Japan, email:
jin@ht.sfc.keio.ac.jp
5 Faculty of Environment and Information Studies, Keio University, Japan,
email: kaz@ht.sfc.keio.ac.jp
6 Faculty of Environment and Information Studies, Keio University, Japan,
email: hxt@ht.sfc.keio.ac.jp We also attach to the chair four pressure sensors. We use FSR406
33
for a pressure sensor and Figure 2 shows how it can be viewed in Confirmation
Sensor Module People Identification
fact. We want to use as least as possible sensors to reduce the cost of User Interaction Module
the system and we think that four is a good number. It is enough for Accelerometer Module
Sensor Learning Module
people recognition issue based on weak evidences.
Inference Result
Pressure
Sensor Normalized
Sensor Data Data
Raw Data
Light Data Receiver Data Processing
Sensor Module Module
Figure 4. The Software Diagram
Figure 2. FSR406 Pressure Sensor
3 Approach
We consider the people identification problem with a small number
of users as a classification problem. The system classifies the data
Figure 3 shows how sensors are placed in the chair. The light sen- from sensors into groups. A group represents a user so the number
sor is used to measure the coverage of user in the chair so it should of groups equal to the number of users exists in the data training set.
be placed in the side of the chair. When a user sit down on the chair, a set of data will be created. The
system will determine which group that this data set belongs to. By
this way, the system can recognize the user who is sitting beforehand.
We used Nearest Neighbor Search Algorithm to resolve this classi-
fication problem. The ”weight” used in the nearest neighbor search
SunSpot algorithm are determined by Softmax Regression Model and Gradi-
( Light Sensor, Accelerometer Sensor ) ent Descent Algorithm. The softmax regression model is discussed
in subsection below.
3.1 Softmax Regression Model
Softmax Regression Model[4] is a supervised learning algorithm
used for multi-class classification. In Softmax Regression Model,
we have a training set {(x(1) , y (1) ), (x(2) , y (2) ), . . . , (x(n) , y (n) )}
of n labeled examples, where the input features are m dimensional
vector x(i) ∈ 0), the cost function
P (y (i) = j|x(i) ; θ) = Pk
θlT x(i)
l=1 e J(θ) is now strictly convex, and is guaranteed to have a unique min-
imize solution. Also, the gradient descent algorithm is guaranteed to
converge to the global minimum. To apply the gradient descent al-
Now, the training is mean finding all θ parameters that minimize gorithm, we also need the partial derivative of this new definition of
the cost function. There are several methods to do it such as gradient J(θ). The partial derivate is shown below.
descent algorithm or limited-memory BFGS algorithm. In this paper,
we used gradient descent algorithm that is described below.
n » –
1 X (i)
∇θj J(θ) = − x (Q(i, j) − P (y (i) = j|x(i) ; θ)) + λθj
n i=1
3.2 Gradient Descent Algorithm
The gradient descent algorithm[5] is a algorithm used to choosing θ
to minimize the cost function J(θ). It starts with some ”initial guess” By using gradient descent algorithm with this equation to mini-
for θ, and that repeatedly change θ to make J(θ) smaller, until hope- mize the cost function J(θ), we will have a working implementation
fully converge to a value of θ that minimizes J(θ). The gradient de- of softmax regression model.
scent algorithm repeatedly performs the update:
θj := θj − α
∂
J(θ) 3.3 Nearest Neighbor Search Algorithm
∂θj
Nearest neighbor search (NNS)[6][7], also known as proximity
search, similarity search or closest point search, is an optimization
This update is simultaneously performed for all value of j. Here, problem for finding closest points in metric spaces. The problem
α is called the learning rate. is: given a set S of points in a metric space M and a query point
To using gradient descent algorithm to minimize the cost function q ∈ M , find the closest point in S to q. In many cases, M is taken to
of softmax regression model, we need to compute the partial deriva- be d-dimensional Euclidean space and distance is measured by Eu-
tive of cost function J(θ). It is shown by the equation below. clidean distance. The Euclidean distance between points p and q is
the length of the line segment connecting them. In Cartesian coor-
dinates, if p = (p1 , p2 , . . . , pd ) and q = (q1 , q2 , . . . , qd ) are two
n » –
1 X (i) (i) (i) points in d-dimensional Euclidean space, then the distance from p to
∇θj J(θ) = − x (Q(i, j) − P (y = j|x ; θ)) q, or from q to p is given by:
n i=1
In particular, ∇θj J(θ) is itself a m dimensional vector, so that its p
l-th element is ∂θ∂jl J(θ), the partial derivative of J(θ) with respect dpq = dqp = (q1 − p1 )2 + (q2 − p2 )2 + . . . + (qd − pd )2
v
to the l-th element of θj . So we can use it to compute the update u d
uX
value of all parameters in softmax regression model. = t (qi − pi )2
But, take a look, if we take each of our parameter vectors θj , and i=1
35
Now, if CP (q) is the closest point to q in set S, we have: dynamic. When a user sit down to the chair, after the system receives
the confirmation from the user, in data training set, the oldest RA of
n
CP (q) = {p|p ∈ S; dpq = min dSi q } this user is replaced by the newest RA. By this way, the system can
i=1
adapt with the change of user’s sitting pattern.
4 Evaluation
3.4 Calculation Process We evaluated this system in two cases. In the first case, we evaluated
with a group of five people and in the other case, we evaluated with
We have discussed about softmax regression model, gradient de- a group of ten people. In both case, one person must sit in a chair
scent algorithm and nearest neighbor search generally in subsections twenty times, ten for training and ten for testing. Figure 5 shows
above. In this subsection, we describe how we use those algorithms the result of first case while Figure 6 shows the result of last one.
in fact. In the case of group of five people, we achieved an accuracy as 90
We use one accelerometer sensor, one light sensor and four pressure percentage and 72 percentage in the case of ten people.
sensors placed on a chair for people recognition, so we have eight
values of sensor data. Accuracy
100
• Ax: The X-axis accelerometer value
• Ay: The Y-axis accelerometer value 80
• Az: The Z-axis accelerometer value
• Light: The light sensor value value
• A1: The first pressure sensor value 60
• A2: The second pressure sensor value
• A3: The third pressure sensor value 40
• A4: The fourth pressure sensor value
When a user sitting down to the chair, an array of 15 records, each 20
has 8 values
0 Person
r = (Ax, Ay, Az, Light, A1, A2, A3, A4) P1 P2 P3 P4 P5
will be created. This array is called RA and it describes the informa-
tion of one sitting time of a user. In our data training set, there are 10
RA for one user. So if the number of user is k, the number of RA in Figure 5. Accuracy for 5 people case
data training set is n = 10k.
RA = {r1 , r2 , . . . , r15 }
We use nearest neighbor search with 8-dimensional Euclidean In the case of five people, there were three people who are iden-
space for this classification problem. But the Euclidean distance tified with the accuracy as 100%. One people with the accuracy as
function we use has a litter different to general function. Because 90%, it means that there was only one mistake.
these eight sensor data affect to result in different ways, we modify With the achievement accuracy as 72% in the case of ten peoples,
the Euclidean distance function like this: this method certainly can be used in a small office or inside a house
v with a small number of users.
u 8
uX
dpq = t θi (qi − pi )2
i=1
5 Related Works
Masafumi Yamada et al.[9] have used 32 pressure sensors placed on
The parameters θ1 , θ2 , . . . , θ8 is the ”weight” of each sensor data and a chair to people recognition. They have tested with a group of eight
we use softmax regression model and gradient descent algorithm to people who are required to sit 20 times, 19 for training and only one
determine them. Our people identify process can be described as fol- for testing. The result is shown in the Figure 7. Their system does
lowing: not recognize user at the time user was sitting down but after few
When a user sit in the chair, a RA is created. We take the aver- seconds, when the values of sensor get steady.The value of sensors
age of all records of this RA and the average of all records of all were collected starting from a few seconds before the user starts sit-
RA in data training set and use softmax regression model to compute ting until the values of the sensor get steady after sitting. From the
the parameters used in nearest neighbor search algorithm.The gradi- data they cut out two parts. One of them is the part during the user is
ent descent algorithm is implemented with learning rate α = 0.001 sitting down, labeled as ”Sitting part” . Another is the part after the
and λ = 0.001 to minimize the cost function in softmax regression sensor value gets steady, labeled as ”Stable part”. The classifier used
model. Finally, we use nearest neighbor algorithm with determined is nearest neighbor method. Every testing data are classified to the
parameters to classify the new RA to one of k class labeled. The re- nearest training data. Used features are classified into four groups to
sult is the user whom this class labeled stand for. There are always investigate how useful the information of pressure sensors is.
10 RA for one user in our data training set, but the data training set is
36
Accuracy also presented the result of experiments which shown that this peo-
ple identification has the accuracy enough to be used in places which
100
have private properties such as inside of a house or a small office.
How to due with other evidences and what is the best way to place
80 sensor to a chair are the things to be discussed in the future. More-
over, the evolution of performance increasing the number of people
need to be studied. We also intend to implement a module to recog-
60 nize the posture of user or the user’s mood.
40
ACKNOWLEDGEMENTS
20 This research was partly supported by National Institute of Informa-
tion and Communications Technology (NICT).
0
P1 P2 P3 P4 P5 P6 P7 P8 P9 P10 Person
REFERENCES
[1] A.K. Jain, L. Hong, S. Pankanti and R. Bolle, ’An Identity Authen-
tication System Using Fingerprints’ , Proc. IEEE, vol. 85, no. 9, pp.
Figure 6. Accuracy for 10 people case 1,365-1,388, 1997.
[2] L. Ma , T. Tan , D. Zhang and Y. Wang, Personal identification based
on iris texture analysis’, IEEE Trans. Pattern Anal. Mach. Intell, vol.
25, p.1519 , 2003.
[3] R. Hill, ’Retina Identification, Biometrics: Personal Identification’ in
Networked Society, A.K. Jain, R. Bolle, and S. Pankanti, eds., Kluwer
• Feature Set1 (FS1): 32 sensors Ω values (32) Academic, pp. 123-141, 1999.
• Feature Set2 (FS2): sum of 32 sensors Ω values (1) [4] C. Do and A. Ng. ’Transfer learning for text classification’ . In Pro-
• Feature Set3 (FS3): time difference of FS1 (32) ceedings of Neural Information Processing Systems (NIPS), 2005
• Feature Set4 (FS4): normalized sensor values of FS3 (1) [5] C. Burges, T. Shaked, E. Renshaw, A. Lazier, M. Deeds, N. Hamil-
ton, and G. Hullender, ’Learning to Rank using Gradient Descent’. In
ICML ’05 Proceedings of the 22nd international conference on Ma-
As we can see, the achieved accuracy in steady part is about 90% but chine learning.
only 56% in the sitting part for average with a group of eight people. [6] Peter N.Yianilos, ’Data structures and algorithms for nearest neighbor
search in general metric spaces’. SODA ’93 Proceedings of the fourth
annual ACM-SIAM Symposium on Discrete algorithms
[7] Stefan Berchtold, Christian Bohm, Daniel A.Keim, Hans-Peter Kriegel,
’A cost model for nearest neighbor search in high-dimensional
data space’. PODS ’97 Proceedings of the sixteenth ACM SIGACT-
SIGMOD-SIGART symposium on Principles of database systems.
[8] http://www.sunspotworld.com, August, 2011
[9] M. Yamada, J. Toyama, M. Kudo, ’Person Recognition by Pressure
Sensors’. Knowledge-Based Intelligent Information and Engineering
Systems, Lecture Notes in Computer Science, Vol. 3684, Rajiv Khosla,
Robert J. Howlett, Lakhmi C. Jain (Eds.), Springer, 2005, 703-708.
[10] Michael R. Bastian, Jacob H. Gunther, Todd K. Moon, ’A simplified
natural gradient learning algorithm’. Journal Advances in Artificial
Neural Systems, Vol. 2011, Article No. 3, Hinwadi Publishing Corp,
Newyork, 2011
Figure 7. Average Recognition Rates by Yamada’s Method
6 Conclusions and Future Works
We have proposed a people identification method based on sitting
patterns of user. This method used weak evidences collected by ac-
celerometer sensor, light sensor, and pressure sensor placed on a
chair to inference who is sitting on it. We considered this problem as
a classification problem and use nearest neighbor search algorithm
with ”weight” to resolve it. The ”weight” used in nearest neighbor
search algorithm is determined by softmax regression model while
the cost function is minimized by gradient descent algorithm. We
37