Columbia University E6893 Big Data Analytics Fall 2015 Final Report

ATC pedestrain tracking

Yan Lu, Mengzhuo Lu, Dingyu Yao

Electrical engineering
Columbia University
yl3406@columbia.edu, ml3806@columbia.edu, dy2307@columbia.edu

Abstract—In this project, we analyzed the 3-D range II. RELATED WORKS
sensor dataset of Asian Pacific Trade Center, a shopping
center in Osaka, Japan. By processing the dataset, we In 2012, a Japanese research team set up 3-D range sensor
analyzed the flow rate in the shopping center using Pig, in ATC shopping mall, they recorded the data collected
and did data visualization to compare the flow rate from 9:40-20:20 throughout the year, and collected in total
under different circumstances. We also did clustering by 92 days data [1]. This dataset has been explored in many
Spark and data visualization for each property in the 3- ways. Zanlungo, Brscic and Kanda researched in
D range sensor dataset, and dug the underlying relations characteristics in group pedestrians, from which we can
between them as well as got many interesting founds. We apply our result on classifying pedestrians into group
also trained a Mahout classifier to distinguish if a shoppers and individual shoppers [2]. This study found
pedestrian walk in a group or individually. Based on group shoppers and individual shoppers different in terms of
these results and explored information, we found some their facing angle and angle of motion, which we can use
invisible fact concealed in boring pedestrian flow data and applied to classify shoppers.
and come up with useful suggestions and strategies. Kanda and Mascot proposed an idea of using robots for
Keywords- Big Data, Pedestrian Tracking, Mall, Sale, Flyer flyer sending in the shopping mall, which our project can
help [3]. Kidokoro and Shiomi researched in people’s
I. INTRODUCTION behavior when encounter a robot in a mall area, which is
done by simulation [4]. A similar topic was introduced by
Asian Pacific Trade Center (ATC), located in Osaka Japan,
is the largest international mall complex in Kansai. In 2013, Hagita, who talked about the potential influence of using
a Japanese research team set up a tracking environment in robots to do public tasks in a city environment [5]. In our
the shopping center. The system consists of multiple 3D project we will propose suggestions and strategies to make
range sensors, covering an area of about 900 m2. The top the robot flyer sending process more effective.
view of the shopping mall is illustrated below. III. SYSTEM OVERVIEW
The datasets we used in this study are "ATC pedestrian
tracking dataset" and "Pedestrian tracking with group
annotations" which were obtained by JST/CREST in japan
[1]. The project primary purposed in enabling mobile social
robots to work in public spaces.
The dataset was collected between October 24, 2012 and
November 29, 2013, Wednesday and Sunday, 9:40-20:20.
The dataset consists of 92 days in total. The data of each day
is provided as CSV files, which each row in a CSV file
Figure 1 Top view of the shopping mall
corresponds to a single tracked person at a single instant,
and it contains the following fields:
The data recorded by these 3D range sensors was made time [ms] (unixtime + milliseconds/1000), person id,
public by the team for research purposes. Our project is position x [mm], position y [mm], position z (height) [mm],
based on these datasets, and intends to distinguish target velocity [mm/s], angle of motion [deg], facing angle [deg]
client and come up with store deploy strategy for ATC
shopping mall and propose strategies for robot flyer senders
and security. We analyzed popular time and area in the
shopping mall; clustered pedestrians based on their
coordinates distribution, and classified them based on their
group behavior.
 Columbia University E6893 Big Data Analytics Fall 2015 Final Report

The third part involves group behavior analysis and

visualization. We firstly used python to pre-process the
dataset, selecting those pedestrians appeared in the group
dataset area. Then appended another column to the end of
each row indicating whether they are involved in a group
Figure 2 Dataset of Single Pedestrian (this information comes from group dataset). We then used
For group data set: Mahout logistic regression algorithm to train and test
models, using datasets from different dates. Then we used
PEDESTRIAN_ID GROUP_SIZE PARTNER_ID_1 ... (list NetworkX to plot the group data, and used Fruchterman-
of ids of all other pedestrians in group) Reingold force-directed algorithm to layout the plot.
NUMBER_OF_INTERACTING_PARTNERS
INTERACTION_PARTNER_ID_1 ... (list of all socially V. EXPERIMENT RESULTS
interacting partners)
i. Pedestrians Group Behaviors Analysis and
Classification
1. Background
ATC shopping mall has thousands of visitors each day, and
they all have different behaviors. We want to distinguish
those behaviors, so we can classify different shoppers, and
use the data for various applications.
As we all know, a major behavior difference among
Figure 3 Dataset of Group Pedestrians shoppers are whether they shop single or with group. Those
two kinds of shoppers tend to buy different categories and
Note, in the two datasets above, the pedestrian ID is the different amounts of products, and their tolerance of
same. Which means if an ID is shown in group dataset, the
interruptions are also different (for example, interruption
pedestrian with same ID in single pedestrian dataset should
can involve being asked to fill in a questionnaire).
also in the dataset.
Therefore, classify shoppers by whether they are in-group or
Because of the time limitation, we considered 42 days (21 not will be beneficial.
Wednesdays and 21 Sundays) with over 123.57GB of data,
From researches we learned that single shoppers and group
and we think this should be enough for the analysis.
shoppers differ in terms of walking angle and looking
IV. ALGORITHM (facing) angle [6]. These information also appears in our
datasets collected by 3D range sensors in ATC shopping
We applied some advanced tools and algorithms we learnt mall. Figure 4 shows us the relations between facing and
from lecture including using Pig in Hadoop to filter, sort walking angles of shoppers in a group of two.
and count data; using Spark to do k-mean clustering and
using Mahout to classify data and make predictions. We
also applied some other tools to virtualize and process data
including Python, Matlab, NetworkX and iNetSoft.
Our analysis can be separated into 3 parts. The first part is
flow rate analysis, in which we used Pig to filter and process
raw data. For instance, since the original dataset sort rows
with time, we used Pig to resort them by person’s ids and
grouped data about a same person. After analysis, we used
iNetSoft to plot our results.
The second part is distribution analysis and visualization.
We firstly used python to pre-processed the dataset and
Figure 4 Angles for a two pedestrians [6]
divide the monitoring area into grids. We also calculated From the
many arguments, such as people intensity using python. We figure we can see that in a two pedestrians group, the two
then used MATLAB to plot in total 42 days of data focusing shoppers walk in the same direction while keeping each
on different characteristics. Pig performed filtering of the
data. We used Spark to do clustering work and clustered the
data based on different categories (for example x, y
coordinates and velocity).
 Columbia University E6893 Big Data Analytics Fall 2015 Final Report

other in his/her eyesight reachable areas. Same rules are

followed by three pedestrians’ systems as well, which is
illustrated in figure 5.
As shown in the figure, the group dataset goes by following
Based on these information, Aanlungo, etc. [2] concludes
rules: the first column lists the id of the currently interested
that that 2 pedestrians, identified as i and j, are socially
person, and the second column shows has/her group size.
interacting while walking towards a common goal, given by
The flowing columns will be the ids of other people in this
a vector 𝑔. If we write their relative positions in polar
group, until another column of a single integer number
coordinates (𝑟, 𝜃), we can write their relative positions as
appears. The rest information in this row is not important
𝑟'( = 𝑟' − 𝑟( . Then angle between this vector and vector 𝑔
for our purpose. They denote who in the group are the
can therefore be written as: interested person are current interacting with.
𝜃 − 𝜋 𝑖𝑓 𝜃 > 0, Note people’s ids in this dataset correspond to the person’s
𝜓= (1)
𝜃 + 𝜋 𝑖𝑓 𝜃 ≤ 0, dataset. Therefore, we used Python to combine the two
As a group, pedestrians will want their common goal in
their vision field, while they also expect their partners in
their vision field [7]. Moreover, they would also want
themselves being in their partner’s vision field [8]. Single
shoppers, however, don’t have such constrains in their
walking and facing angles. Using these information, we can
classify pedestrian’s behavior using their walking and facing
angle data provided by 3D sensors in ATC shopping mall.
2. Analysis of group and person datasets
Since we have found the underline pattern and rules of the
relations between pedestrians’ behavior and their waking
and facing angle data, we would like to analysis the datasets
and give a classifier for future usage. Researchers from the
3D range sensor team manually denoted group information
of pedestrians in part of the shopping center (approximately
in the west most one third of the corridor, close to the Figure 7 Inspecting area for group data
central hall) to part of datasets, and this part of data is also
open for download [9] [2] . This information corresponds datasets together, by appending another column to the end
with the person’s dataset, as introduced before, which of each row of person’s dataset, denoting whether this
means the ids in the group dataset mean the same person in person involves in a group or not. This extra column will be
the person’s dataset of that specific date. The example of the useful in Mahout classification, which will be introduced
group dataset was shown in figure 6. later.
Also, since the group dataset only involved groups in part of
the area in the shopping mall (which is actually area of 𝑥 ∈
−1,9999 and 𝑦 ∈ −1,9999 ), as illustrated in Figure 7.
We filtered the person’s dataset with Python to make sure it
only contains people in this area, which ensures the
accuracy of the classification. As shown by the figure
below.

Figure 6 Group dataset example

 Columbia University E6893 Big Data Analytics Fall 2015 Final Report

Figure 8 Python pre-processed data, with location filtered and

group indicator added

We firstly analyzed the pre-processed datasets using

NetworkX, a big-data visualization tool. We plotted
Figure 11 NetworkX visualization of Sunday data
individual and group datasets on random-selected
Wednesdays and Sundays. The plot was layout using Afte
Fruchterman-Reingold force-directed algorithm. So larger r analyzing, we found the proportion of group pedestrians to
groups will be in the middle, while pairs and unconnected individual pedestrians on Sunday is around 52%, which is
shoppers are on the peripheral. The details can be found in much larger than Wednesday. Also there tends to be more
the figures below. large-groups, but most of the groups are still off two to three
people.
3. Classification by Mahout
To develop a classifier to distinguish groups, we used
Mahout to do classification. We firstly used the Python pre-
processed data for training. W used in total 970MB data for
training, and tested in total 683MB data. We set the column
‘group’ as the classification target, and tried various
predictors, including facing angle, waking angle, velocity, x,
y coordinates and differnet combinations of them, and
finally came up with a best performing classifier.
This classifier uses two predictors, walking angle and facing
angle, both of numeric type. The target variable has 2
Figure 9 NetworkX visualization of Wednesday data categories, indicating if the user involves in a group or not.
The model trained is as follow:
𝑔𝑟𝑜𝑢𝑝 ∼ −12.874×𝐼𝑛𝑡𝑒𝑟𝑐𝑒𝑝𝑡 𝑇𝑒𝑟𝑚
In the figures, red nodes represent those people who + 0.953×𝑤𝑎𝑙𝑘𝑖𝑛𝑔 𝑎𝑛𝑔𝑙𝑒
involved in groups, while blue nodes corresponds to people + 0.791×𝑓𝑎𝑐𝑖𝑛𝑔 𝑎𝑛𝑔𝑙𝑒
whom walk on their own. After analysis we found the
proportion of group shoppers are around 30%. (2)
It can be found that among all of the groups, most of the This model was trained with following parameters: the size
groups have 2 or 3 people. Those nodes on the peripheral of internal feature vector is 20, the number of passes over
are pairs (2-people groups) and individual pedestrians; we input is 100, and the initial learning rate is 50. The average
can refer to this local figure below for details: AUC on the test datasets is around 0.7.
Therefore, using this model we can classify pedestrians and
know if they are walking in groups or not. This information
is useful in many ways, which will be introduced in the
following chapter.
3. Applications
Figure 10 Part of Figure 5 (1) Recommendations on ATC shopping center after group
data analysis and visualization
Same attempt was made on Sunday datasets as well. From
the figure below we can find there are significantly more Based on the datasets analysis in previous chapters, we can
groups than Wednesday. make many recommendations for the restaurants and shops
in ATC shopping mall. For instance, since group shopper’s
 Columbia University E6893 Big Data Analytics Fall 2015 Final Report

proportion is much larger on weekends then it’s on This path is marked black in the figure 5. This is probably
weekdays, the restaurants in ATC shopping center can set because most people on weekdays are just walking by, and
more single tables on weekdays and more couple or larger they tend choose the shortest route (invisible path).
tables on weekends. Therefore, for robots giving flyers as introduced in last
chapter, they can stay on these paths and handing out flyers
Moreover, promotions of lover and family related products
not directly related to in-mall information (because this
(for example flowers, household appliances and children’s
people will not be very likely to shopping). However, on
toys) could also be done more on weekends then weekdays
Sundays, as shown in the bottom picture of figure 5, we can
to make more profits.
see that people are more randomly distributed compared to
(2) Applications of group classifier Wednesday, so more people will be actually shopping then
just walking by. Then sending flyers about in-mall
Individual and group shopper’s behavior is apparently
information (e.g. sale news) will be more effective. And an
different in many ways. And defining groups will be helpful
ideal place for the robots to stay, in this section, is in the
in taking advantage of these differences. For instance, a
black-circled area in the figure below. This area has high
Japanese research group is trying to use robots to send flyers
people intensity and is not likely to cause congestion
[10]. Group shoppers tend to be easier to accept flyer’s from
robots while individual shoppers tends to avoid them [4].
Furthermore, by sending flyers to a shopper group we can
make the flyer sending process more effective. Therefore,
by automatically knowing which person is in groups, we can
make the robot flyers sending more effective.
(2) The classifier can also be used for criminal detection.
For example, if policeman was inspecting a group crime
case, he can use this classifier to quickly filter out all of the
individual pedestrians and it is easier for his to track the
walking path of suspects. Figure 13 Invisible Path & Sale Area
ii. Distribution Virtualization
(because there are spaces on the sides).
1. Pedestrian distribution analysis
There is also one interesting pattern we found: in the narrow
We know from SYSTEM OVERVIEW, that the datasets we corridor in right-down site of the plot, there tend to be more
used contains x, y coordinates of each pedestrian. Using this people on the upper side then down side on Wednesdays,
information, we did visualization for each day in the dataset while on Sundays the intensity are basically the same. This
and plotted the pedestrian distribution. We found many is because on the upper side there are mainly shops, but on
interesting patterns in these plots. the down side there are mainly shop goods display. This
means shops are more attractive then goods display.
Firstly, it is obvious that the people intensity on Sundays is
Therefore, to balance pedestrians on both sides, we suggest
much higher than on Wednesdays. As shown in figure
the shopping mall to move some of the shops from upper
below (in which red denotes high intensity and blue means
side to down side, or set some new shops on the down side.
2．Clustering on pedestrian intensity using Spark
To analysis popular areas in the mall, we conducted
clustering on each day’s dataset using Spark. We use K-
mean algorithm to do the clustering, and each clustering
went through 10 iterations. We divide all the pedestrians in
the mall into 10 clusters and assigned color to them based
on the intensity. Some examples are shown in the figure 14.
It can be clearly seen that Sunday has larger pedestrian
intensity then Wednesday, and most of the clusters lies in
Figure 12 People intensity on Wednesday (left) and Sunday east corridor.
(right)
There is one interesting thing worth noticing. Although the
low intensity). clusters keeps changing, there are some clusters stays in its
Moreover, we found that on Wednesdays, more people tend
to walk in an invisible path, because there are clearly 3
paths in the figure that can be found with large intensity.
 Columbia University E6893 Big Data Analytics Fall 2015 Final Report

place and continues to have low intensity unless special Based on the clustering results, we split the Wednesday
events happens (details will come up in later chapters). We shopping mall into 3 areas: passing by area, shopping area
suggest the shopping mall put effort into these areas, such as and square area.
try to move some popular restaurants or shops there to
balance the pedestrian intensity.
Knowing the locations of cluster centers, we can derive From the figure we can see that the average speed on
many applications. For instance, security services can stay Wednesdays is much higher than it on Sundays. Besides, the
around each center of clustering, so that it distances to all of high-speed area in Wednesdays is larger then that in
people in the cluster will remain controllable. Moreover, Sundays. This could because on Wednesdays there are
this information is also useful for promotions. By putting fewer people, so pedestrians can walk faster and have more
important promotions (including robot flyer senders) on the space to walk. On the other hand, on Sundays the shopping
cluster center of clusters with high intensity will help mall is rather crowded and people cannot walk as fast as
improving the promotion effect. Of course, this strategy will they intended. And the paths spotted earlier for pedestrians
base on the principal of not producing congestion. just passing by has higher speed. So trying to block them in
these paths is not a wise choice for promotions. If we want
3. Pedestrians’ speeds analysis and clustering to send flyers to people on the path (which is tempting
Based on pedestrians’ speeds and their x, y coordinates, we because there are more people there), we should stand on
plotted pedestrians’ speeds comparison between randomly the shopping areas besides the waking path so that we will
selected Wednesdays and Sundays. As shown in the figure not slow the pedestrian flow.
below. 4. Pedestrian Walking Angle Analysis
The last two part provide us with distribution information
from pure location. In this the next 2 section, we will take
velocity, angle of motion and facing angle into
consideration and do advanced data mining.
In this part, we filter out people who are walking by their
velocity. If the velocity is larger than 0.5m/s, we mark them
as walking. And among all walking people, we consider
Figure 15 Pedestrians' speeds plotting their angle of motion and draw the average angle of motion
for people in every grid as shown in figure 17.
However, as we can see from the figure 15, the speed is
linearly related to the color, so it is difficult to distinguish
patterns in the figures. Therefore, we did clustering on
speeds using Spark and plotted the clustering results by
people’s x, y coordinates. The results are shown in the
figure 16 below, where the upper subfigure refers to
Wednesday’s dataset and lower subfigure refers to Sunday's.

Figure 17 Pedestrian Walk Direction

As a result, most people are well behaved, which means

they always walk on left in the narrow areas like east
corridor. The interesting thing is, on Sunday, although the
population of pedestrian increased significantly, people tend
to walk in a more consistent direction. This is probably
 Columbia University E6893 Big Data Analytics Fall 2015 Final Report

because people are not so busy on Sunday, and they can just bar shows the number of person-time shows in the grid for
follow the crowd in relax. This high consistence of motion the whole day.
of angle in each area can not only reduce the possibility of
Comparing with Figure 12, the number of still customer
walking collision (which may cause safety issue), but can
decreases even more significantly on Wednesday comparing
help us find the common sense of pedestrian and further
to Sunday. On Wednesday, there is nearly no person still in
improve our sale strategy. Thus, we can make several
the shopping mall, which verify the conclusion we made
suggestions from this phenomenon:
from Figure 13 that on Wednesday, most people are just
1. People may think the traffic of people becomes massed walking through the shopping mall.
up and may need more security staff, but this is
For the Sunday still customer distribution, we found there
unnecessary. Because on Sunday, people behave more
are some relative popular areas (color in red or orange), and
politely, the possibility of walking collision reduces a
we tried to figure out why these regions tends to be more
lot.
popular, and where should the shopping mall add function
2. For the project of robot sending flayer, on weekends, to make better use of every area.
fix the robot near the south wall of east corridor can be
As shown in Figure 19 (the still customer distribution on
an even better strategy comparing with the strategy we
Sunday, 6th January, 2013), there are in total 8 popular
came up with in figure 13. On the one hand, from figure
regions in the area (marked in black circle).
13, there are still some not so hot areas near the south
wall of east corridor that we can deploy the robot We analysis the layout of the shopping mall on the day, and
without influencing to much of pedestrian’s walking come up with a matching about the possible reason of
comfort. On the other hand, the high consistence on popular areas as shown in Figure 19. To verify our match,
weekend means almost all people go on the left in the we take the facing angle into consideration. As shown in
east corridor, and people walk on the south part of Figure 20, we calculate the average facing angle for still
corridor are people who entering the shopping mall, by customers in a grid for the whole day. The result of Figure
deploy a fixed robot there, we can confirm almost all 20 helps us verifing some matches. The pamphlets and TV
people entering the shopping mall from east corridor area is popular is indeed because of the pamphlets and TV,
can receive the flayer which help intrigue their because the still people there are looking to the wall
shopping interest. direction which has pamphlets and TV on it. Similarly, the
reason why information boards area, shop front area,
5. Still Customer Behavior Analysis
information desk area and bench area are popular is indeed
Compare to the walking pedestrian, still customers tend to because of these functions. Because the facing angles are
spend more time in the shopping mall and have larger indeed facing to the information boards, shop front,
change of consuming in the mall. By analysis their behavior, information desk and sit on the right direction of bench.
we can future filter out the popular areas and detect their While, for the shopping event match, although we cannot
interests. verify it by facing angle, the analysis on the next section
will help us to verify its correctness.
In this part, we also filter out the still customer by velocity.
If the velocity of walk is less than 0.15m/s, we mark the 5. Children Distribution Analysis
person as still. After filtering out the still customer, we plot
the still customer distribution, by the method we used in
part 2 of this section, as shown in Figure 12. In this plot, the

Figure 19 Popular area in Mall

 Columbia University E6893 Big Data Analytics Fall 2015 Final Report

Now, let’s take z axis where processed by python to add an extra column identifier
into consideration. We for time slot (Figure 22), this divide each day into forty time
take people under 1.2 slots.
meter tall (in z axis) as
Then, we use apache pig to process the data to get the
children and filter them
pedestrian flow rate of single day with respect to times. The
out to research their
apache pig is a platform for analyzing large data sets. It
distribution similar to
produces sequences of Map-Reduce programs, for which
what we did in Figure 12.
large-scale parallel implementations already exist, from
As shown in Figure 21,
which we get high efficiency in processing data [11]. The
Figure 20 Still Customer Facing the distribution change of
Angle
procedures are as follow:
children tends to be
larger than all 1) Load data to apache pig source
pedestrians. In other 2) Group the data by identifier times
word, on Wednesday, 3) Count ids in each group of identifier times
there are nearly no 4) Generate data for identifier times vs. number of ids
children in the Therefore, obtain the time vs. number of persons of the
shopping mall, while whole space in mall. Repeat the pig procedures for a
on weekend, the package of datasets of Wednesdays, and then for Sundays
number of children for comparison.
gains significantly.
This actually quite suit
our common sense.
We will further
analysis the detailed
difference for the
number of children in
the next section.
Figure 21 Children Distribution From Figure 21, apart
from the larger fluctuation between weekdays and
weekends, the children distribution data also shows high
concentration on Weekend. Many children spend lots of
time on bench and shop good display area (circled red in
Figure 21). We suggest the shopping mall to assign more
security guard to protect these children and do children topic Figure 23 Comparison of pedestrian flow rate on Wednesday
promotion in these two area. and Sunday; Data includes: Wednesdays (20121024;
20121031; 20121107; 20121114; 20121121); Sundays
iii. Pedestrian Flow Rate Analysis
(20121028; 20121104; 20121111; 20121118; 20121125)
1. Analysis of Pedestrian tracking data The flow density of the whole space (Figure 23) on
A. Weekdays and Holidays flow rate comparison Wednesday increases during morning and remains at a
lower level for most day time. In comparison, on Sundays,
The raw data obtained from origin mobile social robots the trend follows a parabolic pattern, the number of persons
research contains columns of information including time, id in mall increases more rapidly during the morning, peaks
numbers, x_position, y_position, height, velocity, motion during the afternoon from 14:00 to 16:00. The result shows
angle, and facing angle. Note an id number is assigned to in general more pedestrian were tracked on Sundays than on
each person when they first entered the 3D sensor sensed Wednesdays, as also confirmed in our own experience. In
region and continue tracking until the person exit the area addition, we observed the difference in pedestrian density
[1]. Also, The time was measured in million seconds. In between Monday and Wednesday is more apparent in the
order to simplify the plotting load, process the data files afternoon period, roughly 12:00-18:00. This provides clue
for shops in mall to expect excess customers during the
period of times. For example, suggest the shops schedule
more afternoon shift service crews in compensation to
expected increases in demands.

Figure 22 Dataset example with time indentifier after

python code process
 Columbia University E6893 Big Data Analytics Fall 2015 Final Report

From the same set of data, we also plotted the average

number of tracked pedestrian in mall on Wednesday and
Sunday (Figure 24). From which we found the average
pedestrian flow rate on Sunday is about twice as much that
on Wednesday.

Figure 25 Southern stairs area of study

using the stairs is considerably larger on event days only in

a specific time slot, for when the event is holding. On 2012
November 11, the persons using the stairs is flatted all day,
steady constant flow rate through stairs. On 2013 January 6,
there’s an abrupt number of person using the stairs in the
Figure 24 average number of visitors in mall morning from 10:00 to 12:00, in the period the usage of stair
is roughly three times of the regular days. Other than that
the flow rate for the rest of days is similar to regular days.
B. Event day and Nonevent day flow rate comparison
Checking the event schedule in ATC mall for past years
Another approach was made by analysis of the pedestrian [12], we found the Osaka City New Year firefighters’ event
flow density of a corner area for event days and nonevent happened on January 6 in 2013 and the event last two hours
days. From previous research done by D. Brscic, et al. [1], from 10 to 12 in the morning. This confirms our prediction
the use of southern stairs shows distinct differences on event of pedestrian flow rate corresponding to event days. Since
and nonevent days. Excess use of stairs could be a risky we got this information, we can give recommendation on
scenario for public safety that both manager of shopping adding security around the southern stair area for event
mall and customer should have concerned about. In order to period.
evaluate how bad, the use of stairs situation can be for event
days in comparison to regular days, and provide suggestion 3. Applications
for event planer, we perform following analysis. We have found from analyzing pedestrian flow rate,
In this study, we specified the target area of southern stairs shoppers are denser and active on Sundays in comparing on
for study (Figure 25). The stairs lead outside to a wide space Wednesday and the most differences is in the major
in front of building, where sometimes used for different afternoon. Therefore, we can suggest the Shopping mall
events [1]. The enclose southern stairs area is important path have more afternoon shifts service staff on holiday
to the main event area. afternoon, twice as much as than regular days. Moreover,
this can provide useful information for shopper who want to
Referring the vitalization flow rate maps that discusses in avoid peak hours for shopping to shop on weekdays or in
later session, we first picked dataset for dates that with and the morning and night on holidays.
without events. 2012 November 11 and 2013 January 6 are
studied as an example (Figure 26). First, use pig to program When planning an event, it’s also useful to apply the method
filter out the data lines only for enclose area of stairs [11]. we present and discussed above, looking into similar types
Then repeat the process discussed in comparing pedestrian of events that happened before to predict the pedestrian flow
flow rate on Wednesday and Sunday. rate in a target area over a spatial time range. From which
the event organizer gets ideas about how to plan on public
From the visualization plot, observe the values of person space using, tickets selling, assigning security guard, etc. to
guarantee the event’s efficiency and safety.

VI. CONCLUSION
Based on the analysis we did in section V. We find some
interesting phenomenon.
1. There are more people on Weekends but mainly in the
afternoon.
2. Some area is popular only when event happens (which is
a waste of space when there is no event)
 Columbia University E6893 Big Data Analytics Fall 2015 Final Report

3. There is in "invisible path" where more people tends to

walk on it on weekdays. The analysis is very important because it helps the mall to
4. People follows the public order in narrow areas, and find some invisible fact concealed in boring pedestrian flow
people tends to be politer on weekend. data and come up with useful and targeting suggestions and
5. The distribution of children tends to be highly centralized. strategies.
6. The ratio of children and group shoppers on weekend is
higher than weekdays. The program we designed can also be used for other sensor
Based on these interesting found, we came up with some tracking scenario such as traffic tracking, body movement,
suggestions in terms of flyer sending, sale and security etc.
strategy, as well as commercial installation.
Work Distribution: We are a well-operated team, and
Flyer Sending & Sales Strategy everyone make their best effort on the project. Thus
everyone has the same work distribution.
1. Weekday flyer senders (which can be robots) should stay
beside the “invisible path” and advertise business related Acknowledgment
service.
We would like to thank our professor, Ching-Yung Lin for
2. Weekend promotions can happen mainly in the afternoon the support and inspiration for the project. We would also
at the end of the east corridor or shop good display like to thank him for the knowledge and skills he taught
side of corridor to attract shoppers and avoid congestion. throughout the semester. Besides, we appreciate the help
given by our TA, Yongchen Jiang, He offered a lot of
3. Children sales should mainly focus on weekends and only
inspiration and help during the process of the project. We
at shopping good display and bench area.
would also like to thank other TAs, especially Tian Han and
4. Based on the classification result, the mall should Siyuan Zhang for their support.
distinguish the group shopper and do suitable sale and send
suitable flyer targeting these group shoppers, especially for
robot flyer senders.
5. Restaurants in ATC shopping center can set up more
single tables on weekdays, and more couple or larger tables
on weekends. Note that there is no special need for too
many very large table, since most of the groups (both on
weekdays and weekends) are of 2 to 3 people.
6. Promotions targeting couples and families can be done
more on weekends then weekdays.
Security Strategy
1. Setting security at centers of clusters.
2. Do not really need more security staff on weekend.
3. Pay attention to the goods display and bench area to
protect children.
4. Our Mahout classifier can be used for criminal detection,
that policeman can filter out individual pedestrians to solve
group crime case in ATC shopping center.
Commercial Installation
1. Open more stores, while the number of shopping goods
display can be reduced.
2. Open some store or do some sale activity in the triangle
area circled by the “invisible path”.
3. The mall need more staff at information desk especially
on event days and weekend.
4. The mal should come up with more activities to make
better use of the “event area” on non-event days.
 Columbia University E6893 Big Data Analytics Fall 2015 Final Report

APPENDIX group," Physical Review E, vol. 89, no. 1, 2014.

VII. REFERENCES [7] M. L. Knapp, "Nonverbal communication in human
interaction," Cengage Learning, 2012.

[1] D. Brscic, T. Kanda, T. Ikeda and T. Miyashita, [8] K. C. L., "Gaze and eye contact: a research review,"
"Person position and body direction tracking in large Psychological bulletin, vol. 100, no. (1), p. 78, 1986.
public spaces using 3D range sensors," IEEE
[9] D. Brscic and T. Kanda, "Changes in Usage of an
Transactions on Human-Machine Systems, vol. 43, no.
Indoor Public Space: Analysis of One Year of Person
6, pp. 522-534, 2013.
Tracking," IEEE Transactions on Human-Machine
[2] F. Zanlungo, D. Brscic and T. Kanda, "Pedestrian Systems, vol. 45, no. 2, pp. 228-237, 2015.
Group Behaviour Analysis under Different Density
[10] "CREST Research content, Street flyer sending robots'
Conditions," Conference on Pedestrian and Evacuation
design and strategies deployment," [Online]. Available:
Dynamics PED 2014, vol. 2, pp. 149-158, 2014.
http://www.irc.atr.jp/crest2010_HRI/research.html.
[3] A. Sabelli and T. Kanda, "Robovie as a Mascot: A [Accessed 14 December 2015].
Qualitative Study for Long-Term Presence of Robots in
[11] R. D. Schneider, "Hadoop for Dummies," John Wiley
a Shopping Mall," International Journal of Social
& Sons, Inc, 2012. [Online]. Available:
Robotics (IJSR).
www.wiley.com.
[4] H. Kidokoro, T. Kanda, D. Brscic and M. Shiomi,
[12] "2016 Osaka City Fire New Year firefighters' event," 6
"Simulation-based Behavior Planning to Prevent
1 2013. [Online]. Available: http://www.atc-
Congestion of Pedestrians Around a Robot," IEEE
co.com/event/000117/. [Accessed 12 12 2015].
Transaction on Robotics.
[13] D. Brscic, T. Kanda, T. Ikeda and T. Miyashita,
[5] K. Hayashi, M. Shiomi, T. Kanda and N. Hagita, "Are
"Person position and body direction tracking in large
Robots Appropriate for Troublesome and
public spaces using 3D range sensors," IEEE
Communicative Tasks in a City Environment?," IEEE
Transactions on Human-Machine Systems, vol. 43, no.
Trans. on Autonomous Mental Development, vol. 4, no.
6, pp. 522-534, 2013.
2, pp. 150-160, 2012.

[6] F. Zanlungo, T. Ikeda and T. Kanda, "Potential for the

dynamics of pedestrians in a socially interacting