Automatic Detection of Unexpected Accidents Monitoring Conditions in Tunnels

International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395-0056
Volume: 09 Issue: 04 | Apr 2022 www.irjet.net p-ISSN: 2395-0072
© 2022, IRJET | Impact Factor value: 7.529 | ISO 9001:2008 Certified Journal | Page 1167
Automatic Detection of Unexpected Accidents Monitoring Conditions in
Tunnels
Lilesh Deepak Waghmare, Deepa Kulkarni
1,2Department of Computer Engineering, Smt. Kashibai Navale College of Engineering, Pune, India
---------------------------------------------------------------------***---------------------------------------------------------------------
Abstract- As the urban population rises and the number of
motor vehicles increases, traffic pollution is becoming a
major concern in the twenty-first century. Accidents are a
major cause of traffic delays since they not only result in
injuries and losses for those involved, but also in lost and
squandered time for others who are stuck behind the wheel.
The proposed Object Detection and Tracking Technology
(ODTS) would be used and expanded to automatically
identify and control irregular events on CCTVs in tunnels in
conjunction with a well-known deep learning network,
Faster Regional Convolution Neural (Faster R- CNN), for
Object Detection and Traditional Object Tracking. It enables
the detection of a moving target in real time, which is
typically not possible in standard object tracking systems.
The proposed method takes a time frame as input for Object
Detection Bounding Box discoveries, comparingcurrentand
preceding picture bounding boxes to provide a unique ID
number to each moving and detecting object. [3]Avideo clip
is the suggested system. It enables the detection of a moving
target in real time, which is typicallynotpossibleinstandard
object tracking systems. As a result, the computer will
identify any and all injuries. More specifically, because the
training data set is large, it is possible to automatically
improve the ODTS capabilities without modifying the
programme codes.
Keywords- R-Convolutional Neural Network, Object
Detection, Tunnel accident detection.
I. INTRODUCTION
Accidents have been one of the leading causes of death all
over the world. It is difficult to track an isolated tunnel [1]. If
cars outside the tunnel are not aware of the crash, it can
cause further harm. It is usually difficult for policymakers to
understand the complexities of the problem and obtain the
desired assistance. In this paper, we will present an
automated accident warning system.
Object recognition technology has been used effectively to
determine the size and location of target points in
photographs or videos. Several technologies have emerged,
primarily in automotive self-driving, CCTV tracking and
surveillance systems, cancer detection,andsoon.[5]Formal
paraphrase Object detection is another field of image
processing where unique recognition and tracking the
location of known objects over time can beaccomplished. To
monitor objects, however, it is important to first identify the
object type and location in a static image using object
detection. As a result, the effects of object monitoring can be
heavily reliant on the efficiency of the object detection
involved. [3] This object detection system has been
successfully used for tracing targeted pedestrians and
moving cars, crash surveillance in traffic cameras,crimeand
security monitoring in specific local areasofconcern,among
other applications. In a cave,CCTVsurveillanceiscompletely
ineffective. [6] Because the tunnel footage has a poor IL
luminance, the video is heavily affected by the moving
vehicle's tail light or the warning light of the car in service.
The tone of the tunnel video is dark and distinct from the
color of the road outside the tunnel. [2] Formalized
paraphrase for the reasons mentioned above, the video
surveillance device installedonthe roadsoutsidethetunnels
was likely to fail to function properly inside the tunnel. [1]
Formalized paraphrase.
As a result, an attempt is made in this paper to create an
accident detection methodthatcanacquiremovingdetailsof
target objects by integrating an object tracking algorithm
with a deep learning-based object detection mechanism. [5]
Formal paraphrase in the followingpart,thecompleteobject
detection and tracking system (ODTS) procedures will be
outlined in detail. In addition, the tunnel accident warning
method within the scope of ODTS would be considered. This
device detects collisions or unusual occurrences occurring
on moving structures and targets geographic areasonCCTV.
II. LITERATURE SURVEY
On-road vehicle identification is critical for
perceiving driving settings, and localising the observed
vehicle assists drivers in anticipating potential hazards and
avoiding collisions. However, there has been no research on
vehicle identification with partial appearance, and the
process for partially visible vehicle localization has notbeen
investigated. Using stereo vision and geometry, this paper
proposes a novel paradigm for vehicle identification and
localization with partial presence. The initial images from
the stereo camera are then analyzed to create a v-disparity
diagram. Vehicle candidates are created with prior
knowledge of potential vehicle positions on the image after
object detection using v-disparity. Vehicle identification is
completed through deep learning-based verification. A new
partly transparent vehicle tracking algorithm is
implemented for each identified vehicle. This algorithm
senses the vehicle edge on the earth, known as thegrounded
edge, and then selects a reference point for Kalman filter
tracking to map partially visible vehicles. [1]

Author Propose a clear and accurate vehicle
identification system focused on texture and presence
histograms of local vehicles fed into clustering woods. Local
binary pattern-like descriptors are used to remove texture
properties. The align collection of histograms developed by
LBPs spatial for randomly sampled local regions is used to
calculate the dissimilarity between regions of all training
photos. Clustering forests are used to evaluate the match of
histograms. The effectiveness of the proposed approach is
tested on numerous car datasets under various imaging
conditions, and the findings demonstrate that the method
achieves substantial advances over previously reported
approaches. [2]
Smart traffic and information systems necessitate
the processing of traffic data from various sensors in order
to regulate traffic. In this respect, security cameras have
been installed in traffic management and control in recent
years. Several experiments on video surveillance
applications using image recognition methods for traffic
control are being conducted. Video processing of trafficdata
collected from surveillance cameras is an example of an
application for advance warning or data extraction for real-
time vehicle analysis. This paper provides a thorough
analysis of vehicle identificationandrecognitionprocedures,
as well as discussion of various methods for identifying
vehicles in inclement weather. It also addresses thedatasets
used in different experiments to evaluate the proposed
techniques. [3]
The Object Detection and Tracking System (ODTS)
will be implemented and used in conjunction with a well-
known deep learning network, Faster Regional Convolution
Neural Network (Faster R-CNN), for Object Detection and
Conventional Object Tracking, for automated detection and
control of unusual events on CCTVs in tunnels, which are
likely to (1) Wrong-Way Driving (WWD), (2) Stop, (3)
Person out of vehicle in tunnel (4) Fire. This technology
allows you to trace a moving target in real time, which is
unusual in traditional object tracking frameworks. As a
result, the device is capable of detecting all injuries in less
than 10 seconds. The more interesting argument is that as
the training dataset grows in size, the detection capabilityof
ODTS can be automatically improved without any
modifications to the software code. [4]
To hypothesise object positions,cutting-edgeobject
detection networks depend on area proposal algorithms.
SPPnet [1] and Quick R-CNN [2] advancements have
shortened the running time of these detection networks,
exposing region proposal computation as a bottleneck. We
present a Region Proposal Network (RPN) that shares full-
image convolutional features with the detection network,
allowing for nearly cost-free region proposals. An RPN is a
completely convolutional network that predicts object
bounds and objectless scores at each location at the same
time. The RPN is trained from start to finish to producehigh-
quality area proposals, which Quick R-CNN uses for
identification. [5]
The author accomplished the goal in two aspects in
the article. In terms of data processing, we investigated how
to process tracking data effectively by using the parallel
characteristics of iDMA (integrated direct memory access)
and a DSP core; and in terms of data storage, we suggested a
time-sharing approach to address the DSP local memory
(data RAM) usage problem for multiple tracking properties.
In addition, we suggest a new approach for software
architecture that involves two stages of parallel
computations: frame-level parallel computations and
monitoring object-level parallel computations. [6]
Owing to the limited visibility of vehicles in road
tunnels, an accidental crash could quickly be accompanied
by a major secondary accident. As a result, a number of
automatic event monitoring systems have been put in place,
but they have very poor detection rates due tothelowimage
quality on CCTVs in tunnels. To address this limitation, a
deep learning-based tunnel incident detection system was
developed, which demonstrated high detection rates in
November 2017. However, since the object detection
mechanism was limited to still photographs, the movement
path and speed of moving vehicles could not be determined.
[7]
In addition to the RCNNs discussed in this thesis,
there are several other methods for using Convolutional
networks. The identification of model artefacts was
presented as a regression problem. They use a CNN in a
picture window to predict foreground pixels over a coarse
grid for the whole object as well as the top, bottom, left, and
right halves. The projected masks are then converted into
sensed bounding boxes by a grouping mechanism. Szegedy
et al. train their model on PASCALvisual objectclasses(VOC)
2012 training and evaluation from a random initialization
and achieve a mean average precision (mAP)of30.5percent
on the VOC 2007 test. In contrast, an R-CNN of the same
network configuration achieves a mAP of 58.5percent, butit
is pre-trained with supervised ImageNet. [8]
In addition to accuracy, object detection systems
must scale well as the number of object categoriesincreases.
Discriminativelytrainedcomponentbasedmodels(DPM)[8]
are capable of handling thousands of object categories. In
DPM, for example, hash table lookups are used in place of
exact filter convolutions. Their findings indicate that this
technique can operate 10k DPM detectors in around 5
minutes per picture on a desktop workstation. However,
there is a cost. When a large number of DPM detectors
compete, the approximate hashing method results in a
significant loss of detection precision. R-CNNs, on the other
hand, scale very well with the amount of object classes to
detect and almost all processing is spread across all object
groups. [9]
Regardless of scaling actions, an R-CNN on a GPU
will take between 10 and 45 seconds per picture, based on
the network used, since each area is passed across the
network independently. Recent efforts have been made to

minimize preparation and identification time while
increasing accuracy and simplifying the training process.
One of them is Fast RCNN [10], which has better detection
efficiency (mAP) than R-CNN and the other is SPPnet, which
is trained in a single stage using a multi-task failure. SPPnet
training will refresh all network layers, and functioncaching
requires no disc capacity. Another strategy is Faster RCNN.
In this paper, a Region Proposal Network (RPN) is
implemented that shares full-image convolutional features
with the detection network, allowing for virtually cost-free
region proposals. An RPN is a completely convolutional
network that predicts object bounds andobjectlessscoresat
each location at the same time. [10]
III.GAP ANALYSIS
 As there are very few videos of fire in tunnel
accidents. So, the detection of fire is not up to the
mark and requires furthermore learning.
 New models of vehicles sometimes arenot properly
understood so the learning-based algorithm will
need to be made familiar with all newly-launched
vehicles from time to time.
 For processing such unclear and noisy images
various enhancement techniques need to apply.
IV.PROPOSED SYSTEM APPROACH
Fig. System Architecture of Proposed System
The process of defining the architecture, components,
modules, interfaces, and data for a system to meet specified
criteria is known as systems design. Systems design may be
defined as the application of systems theory to the creation
of products. The areas of systems analysis, systems
architecture, and systems engineering have some overlap.
The goal of this system is to provide the user with an
interactive interface that may be used in office settings.
The system architecture encompasses the modules used in
the project as well as the interactions between them based
on data flow and processing. The System is made up of the
following parts:
1. Video Camera
2. Connecting Wire
3. Required Software
4. Fire Brigade
V.WORKING
The system leverages an application to notify local
authorities of any unexpected problems in noisy tunnels.
The raw footage will be captured from the tunnel and
supplied to our system via this approach.Thevideowill have
some frame rate, mostly the frame rate of CCTVcamerasand
regular cameras, which is 25 frames per second.
After achieving the frame rate, it is split up to acquire the
frame rate for 1 second; after obtaining the frame rate for 1
second, the lighting must be adjusted to extract the colour
feature. It should be noted that the image's input size is
around 800*800.
Once the colour feature has been extracted and the lighting
has been modified, the retrieved framewill bepresentedtoa
classifier that has been trained for fire detection.
Whatever colour characteristic we acquire, we must do a
colour transformation. Normally, the picture we get is RGB,
and the colour of the fire in a regular RGB image is yellow.
However, if a yellow cloth or car is already there, identifying
fire will be difficult. To solve this problem, we'll utilise a
colour modification known as YCbCr to extract the Mean,
Median, and Standard Deviation, which will then be used to
identify fire.
The illumination-adjusted picture will then be processed by
a faster CNN. The suggested network will provide a quicker
CNN, which will turn the picture into a block and pass-
through layer and identify the presence of the vehicle. We'll
put the car detection model through its paces. After
detecting the vehicle, a square or rectangular bounding box
and its centroid are obtained. Fromthen,we'll break itdown,
which means that the time x and y-direction, as well as the
centroid, will be recorded. For this, the picture output size is
roughly 300*800.
Because our pictures are two-dimensional, the X-axis will
represent velocity and the Y-axis will represent the vehicles
up and down movement. If the vehicle changes direction, it
will travel uphill or downward. If the two-y vehicle's axis
connects and the colour feature matches fire, a message will
be sent to the fire department and the nearby hospital. We
are using the FRCNN technique for object detection; it
contains two fully connected layers, one after the
convolutional layer and the other aftertheROIpoolinglayer,
so we have two dense layers in total.

Faster RCNN
Figure. Faster RCNN.
Faster R-CNN is a single, unified network for object
detection. The RPN module serves as the ‘attention’ of this
unified network. The R-CNN system trains CNNs from start
to finish to identify the proposal regions as objectorcontext.
R-CNN is mostly used as a classifier and does not estimate
object bounds (except for refining by bounding box
regression). Its precision is determined by the efficiency of
the area proposal module. Several papers have suggested
methods for predicting object bounding boxes using deep
networks. A fully-connected layer is trained in the Over Feat
method to predict the box coordinates for the localization
task that assumes a single entity.Thefully-connectedlayeris
then converted into a convolutional layer for detecting
multiple objects of the same class. The Multi Box methods
produce area proposals from a network in which the last
fully-connected layerpredictsseveral class-agnosticboxesat
the same time, generalising Over Feat's "single-box"fashion.
These class-agnostic boxes are used as R-CNN proposals. In
comparison toourtotallyconvolutional system,theMultiBox
proposal network is extended to a single image crop or
several big image crops (e.g., 224224). The functionality of
the proposal and detection networks are not shared by
MultiBox. We go over Feat and MultiBox in greater detail
later in the sense of our process.
VI. CONCLUTION
We will propose a new ODTS process that combines a deep
learning-based object recognition network and an object
tracking algorithm, and it will demonstrate how complex
knowledge of an object for a particular object type can be
accessed and used. Deep learning trainingsecuredtheobject
detection efficiency of a stable Car object, while Person
demonstrated relatively poor object detectionperformance.
However, in the case of fire, there is a high likelihood of false
identification in untrained videos due to a lack of Fire
artefacts. Nonetheless, by concurrently practising No Fire
artefacts, it is possible to reduce the frequency of false
detections. The deep learning objectdetectionnetwork'sfire
object detection efficiency should be improved later by
securing the Fire image.
VII. FUTURE WORK
This application is simple to build ina varietyofsettings, and
we may add additional functionalityasneeded.Becauseall of
the modules are flexible, reusability is availableasandwhen
needed in this application. Software scope extensibility:This
software improves on extensible ideas such as concealing
data structures, avoiding numerous links or methods,
avoiding case statements on object type, and distinguishing
between public and privateactions.Reuseusability:Itis easy
to upgrade this programme to the next version as and when
needed. Reusablesoftwarelowersdesign,coding,andtesting
costs by spreading work acrossseveral designsandlowering
the amount of code.
We can add a feature to the system that scans the licence
plates of vehicles that are still on the road or have been
damaged, and then searches the database server for the
specific automobile details. A phone or text message would
be sent to the registered owner of the vehicle.
VIII. ACKNOWLEDGEMENT
This work is supported in Lung Cancer Detection System for
medical research and development of any state in India.
Authors are thankful to Faculty of Engineering and
Technology (FET), SavitribaiPhulePuneUniversity,Pune for
providing the facility to carry out the research work.
IX. REFERENCES
[1]. E. S. Lee, W. Choi, D. Kum, “Bird’s eye view localization of
surrounding vehicles: Longitudinal and lateral distance
estimation with partial appearance,” Robotics and
Autonomous Systems, 2021, vol. 112, pp. 178-189.
[2]. L. Cao, Q. Jiang, M. Cheng, C. Wang, “Robust vehicle
detection by combining deep features with exemplar
classification,” Neurocomputing, 2020, vol. 215, pp.225231.
[3]. A. Arinaldi, J. A. Pradana, A. A. Gurusinga, “Detection and
classification of vehicles for traffic video analytics,”Procedia
computer science, 2020, vol. 144, pp. 259268.
[4].K. B. Lee, H. S. Shin, D. G. Kim, “Development of a deep-
learning based automatic tunnel incident detection system
on cctvs,” in Proc. Fourth International Symposium on
Computational Geomechanics, 2019, pp. 140- 141.
[5]. S. Ren, K. He, R. Girshick, J. Sun, “Faster R-CNN: Towards
RealTime Object Detection with RegionProposal Networks,”
in Proc. Neural Information Processing Systems, 2019, pp.
91-99.

[6].A. Bewley, Z. Zongyuan, L. Ott, F. Ramos, B. Upcroft,
“Simple Online and Real-time Tracking,” in Proc. IEEE
International Conference on Image Pro- cessing, 2019, pp.
3464-3468.
[7].K. B. Lee, H. S. Shin, D. G. Kim, “Development of a deep-
learning based automatic tracking of moving vehicles and
incident detection pro- cesses on tunnels,” Korean
Tunnelling and Underground Space Association, 2018, vol.
20, no.6, pp. 1161-1175.
[8] D. Lowe, "Distinctive image features fromscale-invariant
keypoints", International Journal Computer Vision, vol. 60,
no. 2, pp. 91-110, 2017.
[9] A. Krizhevsky, I. Sutskever, G. Hinton, "ImageNet
classification with deep convolutional neural networks",
Proc. Advances in Neural Information Processing Systems,
pp. 1106-1114, 2017.
[10] J. Deng, A. Berg, S. Satheesh, H. Su, A. Khosla, and L. Fei-
Fei. Imagenet large scale visual recognition competition
2012 (ILSVRC2012) [Online]. Available:http://www.image-
net.org/challenges/LSVRC/2012/, 2017.

Automatic Detection of Unexpected Accidents Monitoring Conditions in Tunnels

More Related Content

Similar to Automatic Detection of Unexpected Accidents Monitoring Conditions in Tunnels (20)

More from IRJET Journal (20)

Recently uploaded (20)

Automatic Detection of Unexpected Accidents Monitoring Conditions in Tunnels