Download as docx, pdf, or txt
Download as docx, pdf, or txt
You are on page 1of 3

AUM2015 “We invite abstracts between 500-1000 words with a working title and 50-100 words of

headline summary by 15 March 2015”

Real time station monitoring and modelling for peak flow management. Results of test conducted
on technology deployment at the London Bridge Station and plan for future actions. Commented [SA1]: A few general comments. Please
disregard if you do not think it is appropriate but I
believe we need to change a few things
Authors: Claudio Martani, Simon Stent, Sinan Acikgoz, Vassilis Zachariadis, Steve Denman, Ying Jin, 1)I think we devote too much paper space for
Kenichi Soga Bluetooth. I think it is fair to mention but without
having done the work, it feels unbalanced to devote so
much space to it.
2)I think we can also justify why we don’t need
The high concentration and flow rate of people in stations during rush hour is a prominent risk to
Bluetooth at this stage to realize the vision of people
passenger safety.[maybe cite the recent news articles of crowd problems?] The prediction of passenger flows. All people go to the ticket barriers from
flows in real-time may provide asset managers with the ability to foresee congestions and act early to terminus platforms and that’s really it for the time
avoid itthem, thus improving passenger safety and comfort. being
3) If we are trying to present a vision why don’t we talk
Such passenger flow prediction is critical for the new London Bridge Station where the station layout is about Vassilis’ model more and its potential
constantly changing due to ongoing renovation works. The Cambridge Centre for Smart Infrastructure application to the changing station hoarding details in
and Construction (CSIC) is proposing a method to predict and localize congested ‘hot spots’ in the London Bridge. This could be a good tool to localize hot
spots but also help in determining if the temporary
station in real-time using a micro-simulation model updated by real-time data from an array of
station layout is sufficient for cases where 6 trains
pedestrian counting technologies.
arrive at once, or if temporary works cause closing of a
certain area. I think this grounds the work in a
Our method requires two sources of input data for simulating people flows: (1) a statistical model to particular application which responds to a need rather
describe passenger movements between origin and destination points, based on real-time passenger than a general discussion.
tracking, and (2) a system for real-time people counting at all origin and destination points. Although 4) I think it’d be better if we focused on the London
such inputs are themselvesAlthough a number of technologies exist to gather the aforementioned data Bridge aspect more here. In some parts the discussion
(REF)not unique, the overall approach is novelwe are investigating the suitability of these technologies feels very general, in others specific to London Bridge.
for cheap sensing in a rail station environment. Furthermore, as there are currently, tto the best of our We should choose – if we are focusing on the vision I
don’t think we should discuss particularities of data!
knowledge, our work is one of the first efforts to no approaches which combine the both sources of
data from these sources in real-time order to update inform a forward-looking pedestrian simulation Commented [xx2]: Guys, this is to be decided. I understand
model whichin real-time predicts upcoming flows. that it would be a strong point for the paper. Though
might be a little bit unkind with Costain and NR. Not sure.
To gather the required data, we considered several possible data sourcessensors. Currently, the most What do you think?
common methods for real-time people tracking across large areas use utilize the GSM, WiFi and Commented [SA3R2]: I think this is fine. We are not blaming
Bluetooth signals generated by their mobile devices of the station users. Our initial trial of a GSM- any one, we are only saying it is an important issue.
based tracking system [with CoPath] concluded that, despite the remarkable percentage of pedestrians
tracked (almost 100% versus 20-40% for WiFi and 3-6% for Bluetooth), the technology is ill-suited for
applications in railway stations as the precision of the signal is strongly disturbed by the moving mass
of the trains. On the other hand, the time interval between WiFi signals of available sensors (?) was
considered too long [how long?] for application in stations, that are generally walked across in a few Commented [xx4]: It should be every some 5 min but not
minutes. For these reasons we opted to use Bluetooth tracking that, despite capturing a smaller sure. Plus there is also the issue related to the new policy
proportion of pedestrians, track connected items devices every second and was able tohad the potential of mobile makers to change WiFi ID every few minutes
to define high resolution trajectories. We expect to deploy 9 Bluetooth receivers – provided by Sky that would be nice to cite. Can anyone help here?
High Technology Ltd1 - at the London Bridge station to test the usefulness of this technology for Commented [SA5]: We need to clarify I think! I don’t
building a reliable statistical model of passengers flow in a later stage of the project. Currently, as only understand which interval we are talking about. There is
a few thoroughfare platforms are operational, the flow patterns in the station is limited mostly to entry a wifi company in Cambridge which use WiFi trackers in
and exits. . real-time with a few seconds sampling rate! They don’t
use mobile phones but I think we should be wary of
In terms of real–time people counting, a variety number of technologyies can be adopted. Infra-red making a comment like this. There might be a technology
beam counters and CCTV cameras have both been used extensively in the past, though but these we are unaware of. We can choose
solutions are not optimal for a variety of reasons: beam counters do not provide directional counts, Commented [SA6]: This is work we haven’t done yet! I
CCTV cameras raise privacy concerns, and both can be affected by a lack of precision when the flow would suggest explaining why we haven’t done it after
rate is very high. For this reason at the London Bridge station we have testedinvestigated two talking about it one paragraph. I understand we want to
promising alternative video techniques for people counting. The first uses vertically-mounted Asus deliver the vision.
Xtion Pro cameras, which are low-cost depth sensors that can be used to sense pedestrians by volume
I noticed yesterday that you have to exit the ticket gates
for going f

1
http://skyhightechnology.com/

1
recognition. Their ability to count dense crowds of pedestrians was recently demonstrated in 2. The
second uses vertically-mounted RGB cameras. After preliminary tests at the University of Cambridge, Commented [SA7]: What sort of algorithm does this use?
What tracking do we use Simon? Is there a general
In mid-December 2014 we deployed an array of three depth sensors and an RGB camera at the eastern name/reference?
end of Platforms 12 and 13 of the London Bridge Station in December 2014. The sensors were used to
cross-check the two technologies against each other and test their reliability for counting people. The
sensors and camera were vertically mounted and controlled by a networked mini PC (figure 1, sx). The
location, at the station end of a set oftwo busy train platforms and under a glass ceiling, was
specifically selected for its high footfall and exposure to direct sunlight, which affects the because
performance of depth sensors are known to give a noisy output under direct sunlight. Since the location
would experiences a variety of lighting conditions over thethroughout the course of day and night, it
provided an fair extensive dataset for assessing the reliability of the technology. Software developed by
Counterest3 was used to process depth sensor video data while RGB video data was processed using
software developed within the CSIC.

Results of this first section of tests shows that depth sensors strongly underline the peak times in the
hour 7-8 am and 5-11pm. A sample of one-week data counts for in entrance into and out of platform
area is reported in image 1 (dx).

Commented [SA8]: I think we would also benefit from a


figure of general station layout with sensors.

Figure 1 – camera deployment at the London Bridge Station (sx) with a sample of data recorded in the
week from Monday 2 February to Sunday 8 February, 2015 (dx).

Following up withTo compare with the first remote data gathering, a manual counting were then made
on the records from RGB camera, through a CSIC code script, for four intervals of 10 minutes each
during a day. The four 10 minutes sets of images were chosen within the following time intervals: 8-9
am (high concentration, with sunlight), 12-1pm (low concentration, with sunlight), 6-7pm (high
concentration, low sunlight) and 10-11pm (low concentration, low sunlight).
The manual counting explained the unbalanced ratio between passengers entering in the morning and
these exiting in the evening (see Image 1?). The reason lies in the drop in the depth sensor reliability
under direct sunlight. Indeed the accuracy of the system in place, measured both on precision and recall
of the counting, reveal that sensors are about 90% accurate with no sun, nearly 82% with no direct sun
and about 28% under direct sunshine. Commented [SA9]: Do we need this paragraph before doing
An alternative technology for people counting is to use images recorded from the vertically mounted detailed analysis? Our results seem slightly superficial in
cameras. A tracker model developed by CSIC provides one preliminary source of comparison for the my opinion without graphs to support it. And some
total number of people passing under the designated area. This is potentially an excellent alternative to people might say this is obvious! I

2
Zhang, X., Yan, J., Feng, S., Lei, Z., Yi, D., & Li, S. (2012). Water Filling: Unsupervised
People Counting via Vertical Kinect Sensor. 2012 IEEE Ninth International Conference on Advanced
Video and Signal-Based Surveillance (AVSS) (pp. 215-220). Beijing: IEEE.
3
http://counterest.net/en/

2
be used in locations exposed to the sunshine, where depth sensors have a lower accuracy. Commented [SA10]: I think we can delete this paragraph it
Based on results of test conducted so far it seems reasonable to imagine that a combination of depth repeats what was said before.
sensors and vertically mounted cameras for people counting and of Bluetooth receivers for passengers
tracking will provide reliable input data for simulating passengers flows in dense transportation hubs. Commented [SA11]: I think we need more concrete results
section. I think we should say that this project outlines
the advantages/disadvantages of different people
counting techniques, and combining this with
microsimulation, it seeks to investigate passenger flows
in temporary station layouts. On the basis of this result,
we believe that the vision outlined earlier has significant
potential to transform station management with early
prediction of crowds using ubiquitous and cheap sensing
technologies.

You might also like