You are on page 1of 13

An Automated Approach to Prevent

Suicide in Metro Stations

Anindya Mukherjee and Bavrabi Ghosh

Abstract Every year, we find a lot of people committing suicide in metro. To


avoid such incidents, we have designed a system that would take the captured video
of preinstalled CCTV cameras in metro stations and analyze them. On breaking
down the captured video into frames, the region of interest for each will be cal-
culated and their histogram found. Processing the histogram values, if danger is
detected then an alert message will be fired. In future with this triggering of the
message a physical barrier, installed beforehand at the edge of the platform, will
come up thereby preventing the victim from making the suicide attempt, saving
his/her life and hours of harassment for others.

Keywords Histogram ⋅ Trigger message ⋅ Suicide attempt ⋅ CCTV footage

1 Introduction

Since its inception, the metro rail service has served as a viable, cost-effective, and
convenient mode of transportation. Yet, there is an issue that diminishes the certain
mentioned glow. That issue is the rampant rate of suicide. According to the journal
by Roychowdhury et al. [1], a total of 58 suicide cases have been reported since
1987 till 2008. Suicide not only takes away a person’s precious life but also results
in hours of delay leading to unnecessary harassment for those who rely on the
uninterrupted service. Now, the death in some cases is not even caused by physical
injury inflicted by the moving train, even after applying brakes the person may get
electrocuted due to the high current in the third rail.

A. Mukherjee (✉) ⋅ B. Ghosh


Department of Computer Science and Engineering,
Institute of Engineering & Management, Kolkata 700091, India
e-mail: anindya.pg@gmail.com
B. Ghosh
e-mail: bavrabi@gmail.com

© Springer Nature Singapore Pte Ltd. 2017 753


S.C. Satapathy et al. (eds.), Proceedings of the 5th International Conference on Frontiers
in Intelligent Computing: Theory and Applications, Advances in Intelligent Systems
and Computing 515, DOI 10.1007/978-981-10-3153-3_75
754 A. Mukherjee and B. Ghosh

Statistics show that most of the suicide attempts take place in the busy office
hours, disrupting the day to day life of passengers, causing a lot of trouble. The
journal by Roychowdhury et al. [1] reported an average of 15 deaths between 1 pm
and 3 pm followed by an average of 13 deaths between 3 pm and 5 pm, while the
death toll was comparatively less at other times of the day.
A measure to prevent all these unfortunate deaths can be taken by adopting our
system. We plan to capture the live feed of existing CCTV (Closed-circuit televi-
sion) cameras in metro rail stations and apply our algorithm to that.
We split the CCTV camera feed (in the form of video) into a series of frames.
We choose the region of interest from each frame and do further processing. During
further processing, we calculate the histogram of consecutive frames and find their
difference. If the difference lies in a particular range, we identify it to be a danger
situation, during which someone is trying to commit suicide. In this situation, we
trigger the alert message. If the histogram difference does not lie in the specified
range, then the system goes on without any interruption.
The different sections of this paper are organized as follows: the “Literature
Review” section contains the summary of research works related to our topic or
from where we have got hints about how to proceed with our work and what the
possible ways of obtaining a certain result are. It also has discussions about how our
designed system is different from others and why we cannot implement the dis-
cussed methods in our systems because of presence of a few constraints. The
“Proposed Method” section contains a detailed description of how the system has
been implemented, with pictorial representation. Following this section, we have
the “Result and Discussion” part where the ultimate result obtained by the system is
discussed in details with an overview of what the ultimate goal of the project is. The
“Future Scope” section has further proposals about the hardware implementation of
the project, followed by the “Conclusion” section that gives an insight of how our
project can benefit society. The “References” section holds record of all the
materials we referred to, for our work.

2 Literature Review

We have come across a few research works in the same field comprising various
other methods to detect unusual behavior or prevent suicide. The outcome of our
analysis of these systems is given below:
1. Zhong [2] developed an unsupervised technique for detecting unusual events in
a large video set. In this technique, the goal was to detect unusual activity in
certain scenarios, like patient monitoring in a hospital or cheating in a poker
game. The researcher has made a few prototypes of the probable case that can
take place and then matches them with the video he/she is capturing of the event
taking place in front. For example, in case of a poker game the different types of
moves are recorded and stored as a prototype. Now, when the actual game is
An Automated Approach to Prevent Suicide … 755

going on, the whole match is recorded through a camera. If it is a clean game,
then there would be huge amount of similarity between the actual video and the
prototype of a fair game. If any dissimilarity is found, then unusual activity is
detected and in case of a poker game, it has to be some foul move for cheating.
Same applies for patient monitoring in a hospital.
2. Shim et al. [3] proposed a model of suicide prevention along one side of a bridge
which has been based on region of interest. Through the video footage captured
by CCTV cameras, the region of interest (target danger zone) is identified and
marked beforehand. Now, the images (frames) are converted to binary form and
blob analysis of the frames is done. Here, the background image, that is, the
river on which the bridge is situated, is constant. So, whenever anyone tries to
commit suicide by climbing the railing of the bridge, a percentage of the per-
son’s binary image’s blob is calculated to be present in the danger zone. If the
percentage is higher than allowed threshold value, then an alarm is raised.
3. Rodriguez et al. [4] designed a crowd detection technique where optimization of
joint energy function, combining crowd density function, is used to track the
number of people in a crowd.
4. Ali Dağlı et al. [5] took into consideration the presence of security cameras on
metro stations and that the staff cannot monitor all camera feeds simultaneously.
The detection of unusual events and generation of warnings to the staff is their
main goal. This paper was published in April 2013, yet, there have been mul-
tiple incidents of suicide even after using this approach.
Studying the above-mentioned works and models, we have come across a few
components which are not applicable in our project. These include:
1. Like the model by Zhong et al. [2] to detect unusual activity, maintaining a
prototype for this project is next to impossible as it will result in huge amount of
memory usage and file usage. Also matching a live video with prototype will
need some time, which will not make the project efficient and here, we need
utmost efficiency as being late by a fraction of a second, can cause a life to end.
2. In our project, the maximum number of suicides occurs when the train is
entering the station. So the background is not constant. Therefore, doing a blob
analysis in a changing background is not a wise decision for our project as our
region of interest is the tunnel through which the train is traveling. This is in
contrast with the model proposed by Shim et al. [3].
3. We cannot follow the crowd density function used in the technique by Rodri-
guez et al. [4] as we plan to take into account the edge of the platform and the
position of the camera which would not capture major part of the crowd.
Therefore, following this method would not help much in our system.
4. In our paper, the method that we are following is simpler and easy to implement
compared to the approach by Ali Dağlı et al. [5]. We are not specifically
detecting the train when it is entering the platform, neither does the system have
to detect any yellow line intrusion. Thus, computational complexity reduces and
756 A. Mukherjee and B. Ghosh

it is better when one is going to implement it in real time because fast processing
is of high importance in life and death situations.

3 Proposed Method

There are roughly four major steps in which we have carried out the work to
prevent suicide in metro. First step is to connect our system to the CCTV cameras
installed in the metro rail platforms and read the video stream. This video stream is
read at the rate of 25 frames per second, in the video format ‘RGB24’.
Following this, in the second step, the frames from the video are extracted. This
is done in such a way that, at any time, we have the current frame and can also
cross-reference with a subset of the elapsed frames. We have stored the current
frame and its preceding frame in two separate variables. These variables contain
information about the spatial distribution of pixel intensity values in raster graphics
format, for consecutive frames. We retrieve these frames using a loop that ends with
the stream (Figs. 1, 2 and 3).
In the third step, we perform different image processing techniques on the
extracted frames (Figs. 1 and 4), within the body of the loop. Since the CCTV
cameras in metro stations have fixed fields of view, the location of the tunnel does
not change with each frame. Hence, select the tunnel from the entire scene by
specifying coordinates to represent the boundary of the tunnel within the scene.
Now, using the coordinates, we apply polygonal region of interest by calling the
“roipoly” function and thus extract the region of interest (Figs. 2 and 5) from each

Fig. 1 Original preceding


frame
An Automated Approach to Prevent Suicide … 757

Fig. 2 Extracting region of


interest of preceding frame

Fig. 3 Result after extraction


(preceding frame)

Fig. 4 Original next frame


758 A. Mukherjee and B. Ghosh

frame, storing the results of consecutive frames (Figs. 3 and 6) in two different
variables. The tunnel is taken as the region of interest for better detection while
reducing time complexity (Figs. 7, 8 and 9).

Fig. 5 Extracting region of


interest of next frame

Fig. 6 Result after extraction


(next frame)

Fig. 7 Region of interest of


preceding frame
An Automated Approach to Prevent Suicide … 759

In the fourth step, using the previously stored frames, following the extraction of
region of interest, we calculate histogram of consecutive frames. These histogram
values (Figs. 8 and 10) are now stored for further comparison.
In the fifth and final step, we calculate the difference in histogram values for the
consecutive frames. When there is no train in the tunnel then its histogram does not
change much. Also, when the train enters the platform, the changes in the histogram
take place at a certain rate. In normal situations, when no one is trying to commit
suicide or no foreign object enters the tunnel, then the histogram value either lies
below 0.0037 or above 0.0040. However, in both situations, if a foreign body enters
the frame, the histogram change is noticeable. When a person tries to jump in front

Fig. 8 Histogram of the


preceding frame

Fig. 9 Region of interest of


next frame
760 A. Mukherjee and B. Ghosh

Fig. 10 Histogram of the


next frame

of the train, then the histogram value lies within the range of 0.0037–0.0040.
Hence, when the histogram difference lies in this range, we detect an unusual
scenario in the tunnel. On detection of the unusual scenario, an alert message is
fired (Fig. 11) thereby alerting the authority about the intrusion in metro service.
Seeing this message, the security personnel can save the person.
The algorithm used by our system and the flow of events through which our
system passes on following this algorithm, are given on the next page.

Fig. 11 This is the alert


message generated when a
suicide attempt by any person
is detected
An Automated Approach to Prevent Suicide … 761

3.1 Algorithm
762 A. Mukherjee and B. Ghosh

3.2 Flow of Events


An Automated Approach to Prevent Suicide … 763

4 Result and Discussion

During execution of our system, we have used as input, recorded CCTV footage
from a metro station in the form of a video file where it is seen that a person
commits suicide. Our system can take video footage from CCTV cameras without
needing any enhancement and it also accepts video files in Audio Video Interleaved
format. If the test video is not in this format, preprocessing is needed to convert it.
We have performed our experiments on an Intel® Core™ i5-3230 M machine with
4 GB of RAM, running Windows 10 64-bit.
For the result, we have designed a GUI (Graphical User Interface) window
where we have made two panels. In one panel we run the video without applying
the algorithm. The train passes over the person and she dies (Fig. 12). We find the
person was able to commit suicide without any intervention when the algorithm is
not applied.
Upon applying the algorithm, we find that as soon as the person’s leg enters the
region of interest (the tunnel) the system detects a danger situation based on his-
togram difference and an alert message is fired (Fig. 13).There is very little chance
that someone will stoop to confirm the arrival of the train as the sound is enough.
Hence, the possibility of generation of false alarm is very feeble. Also, if this system
is installed then passengers will be told beforehand about the rules and regulations
to be followed on the platform when the train has not yet arrived.
The number of experiments performed by us is 30, including both situations
where suicide does not occur and also situations where suicide occurs. We found
that the alert is fired only in the latter case. Thus, the success rate is 100%. Time

Fig. 12 Result without algorithm applied


764 A. Mukherjee and B. Ghosh

Fig. 13 Result with algorithm applied

taken to simply display 60 frames was found to be 4.21 s and time taken after
processing using our algorithm was found to be 7.95 s. Hence the lag, per frame
was found to be 0.062 s using the formula (1).

lag = ðtime with algorithm − time without algorithmÞ ̸frames . ð1Þ

The time required to generate the graphical user interface on the screen was
found to be 0.027 s on the first run, and the time needed to display the message
upon detection of suicide activity was faster than what could be measured.

5 Future Work

Till now, the system that we have developed can display an alert message on
detection of a person trying to commit suicide in metro. That alone would not help
much. In future, we plan to implement a physical barrier (Fig. 14) that will stop any
person from falling or jumping on the track, using thin fiber sheets. These sheets
will be installed at platform edge with sponge padding to prevent any injury. This
will further automate the system by directly saving life the moment a suicide
attempt is detected.
An Automated Approach to Prevent Suicide … 765

Fig. 14 Physical barrier


comes up preventing the
person from jumping onto the
track

6 Conclusion

Suicide is the extreme form of decision a depressed mind takes when he/she feels
there is no more hope left for them. Our system aims to save the lives of all these
people, saving their families from grievance and also ensuring that co-passengers
do not face any inconvenience. Hence, this system can be beneficial to many
people.

References

1. U.B. Roychowdhury, M. Pal, B. Sukul: Suicides in Kolkata metro railway. In: Journal of
Indian Academy of Forensic Medicine, 2009. vol.: 31, Issue: 2, pp. 118–121.
2. Hua Zhong: Detecting Unusual Activity in Video. In: Computer Vision and Pattern
Recognition, 2004. CVPR 2004. Proceedings of the 2004 IEEE Computer Society Conference,
Page(s):II-819 - II-826 Vol.2, DOI:10.1109/CVPR.2004.1315249.
3. Young-Bin Shim, Hwa-Jin Park, Yong-Ik Yoon.: A Study on Surveillance System of Object’s
Abnormal Behavior by Blob Composition Analysis. In: International Journal of Security and
Its Applications Vol.8, No.2 (2014), pp. 333–340.
4. Mikel Rodriguez, Ivan Laptev, Josef Sivic, Jean-Yves Audibert.: Density-aware person
detection and tracking in crowds by Blob Composition Analysis. In: 2011 International
Conference on Computer Vision, pp. 2423–2430, DOI:10.1109/ICCV.2011.6126526.
5. Mehmet Ali Dağlı, ÇiğdemEroğluErdem: Auto-detection of unusual events in metro stations
via security cameras. In: Signal Processing and Communications Applications Conference
(SIU), 2013 21st, pp. 1–4,DOI:10.1109/SIU.2013.6531414.

You might also like