You are on page 1of 6

Proceedings of the 2014 IEEE

International Conference on Robotics and Biomimetics


December 5-10, 2014, Bali, Indonesia

Development of a Person Counting System Using a 3D Laser Scanner


Shun-ichi Akamatsu1 , Naohiro Shimaji2 and Tetsuo Tomizawa1

Abstract— The purpose of the present study is to develop


a person counting system using a 3D laser scanner. The
system consists of a person-detection component, a tracking
component, and a counting component. The person-detection
component detects human bodies as a point cloud, even if a
number of people are close to each other. To achieve this,
we use a method of grouping the point cloud towards the
bottom from the top. The tracking component tracks the
detected person using a Kalman filter. The counting component
counts the number of people who enter and exit a structure.
Experimental results obtained in the present study reveal that
the proposed system can recognize individual pedestrians and
count the number of people passing through crowded outdoor
environments.

I. INTRODUCTION
Fig. 1. Target environment
It is very important for administrators of public facilities
or shops to know the number of visitors using their facilities.
In order to determine the exact number of people that visit
a particular establishment, the number of people passing pass through a doorway, so that the door-opening timing
through the doorways of the building must be counted. is appropriate. The 3D laser scanner used in the system is
Doorways located in semi-outdoor environments are affected robust to illumination changes. In addition, it is not necessary
by rain and sunlight. In addition, occlusions often occur to install a pole or a sensor on the floor or wall because
in crowded passages. Therefore, existing systems have poor the sensor is attached to the top of the door. Moreover, this
accuracy. system has no dead space and does not adversely affect the
There are conventional methods by which to detect a landscape.
person in such an environment. One typical method for hu- In the present paper, we propose a method by which to
man detection involves the use of ceiling cameras [1][2][3]. count the number of people entering and exiting a structure
K. Terada et al. and D. Beymer used a stereo camera attached using a 3D laser scanner that was developed for use as an
to the ceiling [1][2]. S. Velipasalar et al. used a single automatic door sensor.
camera attached to the ceiling [3] for human detection. These
systems often fail to detect a person when changes in the II. S ENSOR
lighting conditions occur. Methods using 2D laser scanners In the present study, we used the 3D laser scanner de-
have also been proposed [4][5]. K. Nakamura et al. used veloped by Hokuyo Automatic Co., Ltd. as a door sensor.
multiple single-row laser scanners installed on a pole inside a The measurement is based on a time-of-flight (TOF) method,
passageway [4]. K. Katabira et al. are also using a single-row in which the distance is calculated from the time difference
laser scanner attached to the ceiling [5]. A 2D laser scanner between emitting a laser beam and receiving the reflected
can be used to measure distance. However, blind areas often beam. The specifications of the sensor are shown in Table I.
occur in crowded environments. In addition, sensors must be The detection area on the ground is 5.0 m in width and 2.8
installed at the same height as the flow line of the people, m in depth when the sensor is attached at a height of 3.1 m.
which may not be acceptable for building owners because of The scanning area is shown in Fig. 2. The data sent from the
potential problems associated with landscaping. sensor includes the distance to the object and the projection
Recently, we have been developing an intelligent au- angle of the laser beam, and it is possible to obtain data
tomatic door system using a 3D laser scanner [6]. The (5,440 points/frame) at 10 Hz.
system measures the velocity, the position, and the width
of a pedestrian and judges whether the person intends to III. A LGORITHM
1 S. Akamatsu and T. Tomizawa are with the Department of Human In this section, we describe the algorithm for counting the
Media Systems at Graduate School of Information Systems, The Univer- number of people passing through a doorway. The algorithm
sity of Electro-Communications, Choufu, Tokyo, Japan, {akamatsu, can be roughly divided into person-detection, tracking, and
tomys}@taka.is.uec.ac.jp
2 N. Shimaji is with the Department of Engineering Hokuyo Automatic counting components. First, the sensor data is separated into
Co., Ltd., Osaka, Japan shimaji@hokuyo-aut.jp individual objects by grouping. Next, the individual objects

978-1-4799-7397-2/14/$31.00 © 2014 IEEE 1983


>ĂƐĞƌ^ĐĂŶŶĞƌ

^ĐĂŶWŽŝŶƚƐ

Fig. 2. Overview of the sensor and scanning area

TABLE I
S PECIFICATIONS OF THE SENSOR (a) Photograph of the experiment
Specification
Optical source laser diode
Type of Measurement TOF (pulse-modulated signals)
Scanning Device Resonant mirror
Horizontal Range 72 deg
Vertical Range 42 deg
Frame Rate 10 Hz
Number of Observation Points 5,440 points/frame
Temperature Resistance -20 to 50◦ C
Size 127 (H)× 230 (L)× 83 (W) mm

are identified as being/not being people, and the objects


identified as people are tracked. Finally, the system counts
the number of people who enter and exit a structure based
on the movement trajectory of the tracked individuals.

A. Person detection
The person-detection component extracts human figures
from point clouds observed by the 3D sensor.
1) Grouping: The point clouds consists of more than one
person and other moving objects.
(b) Raw point cloud
If these objects are sufficiently separated, it is possible
to identify each person using a simple grouping method
Fig. 3. Raw point cloud data
based only on the point-to-point distance. However, if the
distance between individuals is very short, persons are cannot
be separated by the simple grouping method. On the other
hand, since the head of a walking person is generally not in a horizontal distance of less than 0.2 m (x and y axes)
contact with another object, we adopt a method that groups between a point and a previously labeled point and a
the point cloud from top to bottom based on the heads vertical distance of less than 0.5 m (z-axis distance). If
of individual people, which makes it possible to separate there is no labeled point, then the second highest point
the point cloud associated with each person. This grouping is assigned a new label, i.e., group B.
method is described below. 4) This procedure is repeated until all points in the cloud
are labeled.
1) The point cloud is sorted in ascending order according
to the heights of the points (on the z-axis). The raw point cloud data is shown in Fig. 3. The results
2) The highest point is labeled group A. of the grouping are shown in Fig. 4. Points of the same color
3) Next, focus on the second highest point. If the distance are assigned to a single object.
between the point and a point of a previously defined 2) Determination of whether a detected object is a person:
group is short, the same label is assigned to the point of We next consider whether or not a detected object is a person.
focus. The thresholds for determining the closeness are The determination is based on the whole-body size of each

1984
TABLE II
PARAMETERS AND CONDITIONS FOR PERSON DETERMINATION
Parameter Parameter name Condition
Range of the height of the object objectheight > 0.2 m
Area of the object objectarea > 0.05 m2
Width of the object objectwidth > 0.2 m
Depth of the object objectdepth > 0.2 m
Range of the height of the head headheight > 0.1 m
Area of the head headarea > 0.02 m2
Width of the head headwidth > 0.1 m
Depth of the head headdepth > 0.1 m
Ground height of the head headz > 0.5 m

Fig. 4. Result of grouping

Fig. 6. Result of person detection (red rectangle)

• objectarea
1) Project the point cloud of the object onto the x-y
plane.
2) Calculate the smallest convex polygon that con-
tains all point clouds projected onto the x-y plane.
3) Calculate the area of the convex polygon. Here,
Fig. 5. Extracted head points the area is defined as objectarea .
• objectwidth , objectdepth

1) Calculate the smallest rectangle that contains


group and the shapes of the upper 10-cm part of each group, the convex polygon, which is calculated when
which are considered to belong to the heads of individuals. objectarea is calculated.
The body parameters consist of range of height, area, width, 2) The long side of the rectangle is defined as
and depth. The head parameters consist of range of height, objectwidth , and the short side of the rectangle
area, width, depth, and ground height. The results of head is defined as objectdepth .
detection for each object are shown in Fig. 5. The object is The parameters for head of the object are calculated using
determined to be a person if all of the parameters satisfy the the same method.
conditions shown in Table II. The parameters are set to be
low so that the system does not generate false negatives. The B. Tracking
method used to calculate the parameters is shown below, and In the present paper, we assume people undertake uniform,
the results of determination are shown in Fig. 6. linear motion. Therefore, the system tracks people using
• objectheight a Kalman filter. If the center of the person is defined as
Calculate the minimum and maximum heights of the center of the detected object described in Section III-A,
the point cloud of an object ( zmin , zmax ). Here, the center of the person may vary significantly because the
objectheight is defined as (zmax − zmin ). size of the detected object varies according to whether arms

1985
Fig. 7. Result of person tracking (blue lines)

(a) Overview of the sensor and doorway

ŽƵŶƚůŝŶĞ

Fig. 8. Count line

or baggage are also detected in the same group. However,


the center of the head of a person is not affected by such
variations. Therefore, in the present paper, the center of a
person tracked using a Kalman filter is defined as the center (b) Photograph of the experimental setup
of the person’s head. The tracking condition ends when a
person who is being tracked exits the effective sensor area or Fig. 9. Experimental environment
cannot be detected for more than half a second. The tracking
results are shown in Fig. 7.
line was defined as 0.9 m from the door. We then counted
C. Counting the number of persons who enter or exit the structure.
The count line shown in Fig. 8 defines the effective area The experiment was run for approximately 90 minutes.
of the sensor used to count the number of people entering The total number of persons entering and exiting during the
and exiting a structure. The person tracked in Section III- experiment was 361 and 388, respectively. Scenes (a), (b),
B is determined that the person exists in front of the count and (c), taken during the experiment, are shown in Figs. 10,
line or behind it. The system determines that the person is 11, and 12. Figure 10 shows the point cloud and counting
entering the structure if the person moves from the front of results and a photograph for a situation in which people are
the count line to behind it. Moreover, the system determines simultaneously entering and exiting the structure. Figure 11
that the person is exiting the structure if the person moves shows the point cloud and counting results and a photograph
from behind the count line to in front of it. for a situation in which people are close to each other. Figure
12 the point cloud and counting results and a photograph for
IV. E XPERIMENT a situation in which four people are simultaneously entering
We attached the 3D laser senor at a height of 3.1 m at the and exiting the structure. It was possible to detect persons
top of the entrance of a store, as shown in Fig. 9. The count using the proposed algorithm in Scenes (a), (b), and (c).

1986
(a) Point cloud and results of counting
(a) Point cloud and results of counting

(b) Photograph captured from the view-


point of the sensor (b) Photograph captured from the view-
point of the sensor

Fig. 10. Experimental scene (a)


Fig. 11. Experimental scene (b)

The total number of people entering and exiting the structure


as determined by the proposed system was 365 and 389,
respectively. Therefore, the error in the proposed system was
less than 2%.
V. C ONCLUSIONS
In the present paper, we proposed a method by which to
count the number of persons entering or exiting a structure
using a 3D laser scanner. We focus attention on a person’s
head, the grouping method was based on the height of
the point cloud, so that the system was able to detect
multiple individuals in close proximity to one another. A
demonstration experiment was conducted in order to confirm
the possibility of counting people in various situations,
such as two people in close proximity entering a structure (a) Point cloud and results of counting
and multiple people simultaneously entering or exiting the
structure. The proposed system can recognize individual
pedestrians and can count the number of people passing
through crowded outdoor environments.
Further experiments are necessary in order to confirm
the detection accuracy, for example, for differences in body
height, as in the case of children and adults, or for people
carrying baggage or wearing hats.

(b) Photograph captured from the view-


point of the sensor

Fig. 12. Experimental scene (c)

1987
R EFERENCES
[1] K. Terada, D. Yoshida, S. Oe, and J. Yamaguchi, “A method of
counting the passing people by using the stereo images”, International
Conference on Image Processing, vol. 2, pp. 338-342, 1999.
[2] D. Beymer, “Person counting using stereo”, Workshop on Human
Motion, pp. 127, 133, 2000.
[3] S. Velipasalar, Y.-L. Tian, and A. Hampapur, “Automatic counting
of interacting people by using a single uncalibrated camera”, IEEE
International Conference on Multimedia and Expo, pp. 1265-1268,
2006.
[4] K. Nakamura, H. Zhao, X. Shao, and R. Shibasaki, “Human Sensing
in Crowd Using Laser Scanners”, Laser Scanner Technology, Dr. J.
Apolinar Munoz Rodriguez (Ed.), ISBN: 978-953-51-0280-9, InTech.
[5] K. Katabira, K. Nakamura, H. Zhao, and R. Shibasaki, “A method
for counting pedestrians using a laser range scanner”, In: 25th Asian
Conference on Remote Sensing (ACRS 2004), Thailand, November
22-26, 2004.
[6] D. Nishida, K. Tsuzura, S. Kudoh, K. Takai, T. Momodori, N. Asada,
T. Mori, T. Suehiro, and T. Tomizawa, “Development of Intelligent
Automatic Door System”, Proceedings of the 2014 IEEE International
Conference on Robotics and Automation, pp. 6368-6374, 2014-06-04.

1988

You might also like