See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/384165020

Artificial Intelligence based System to Prevent Animal Accidents in the Railway

Tracks

Conference Paper · July 2024

DOI: 10.1109/ICISC62624.2024.00029

CITATIONS READS
0 50

4 authors, including:

R.Bagavathi Lakshmi Dr M Sakthivanitha

Vels University Vels University
14 PUBLICATIONS 25 CITATIONS 11 PUBLICATIONS 37 CITATIONS

SEE PROFILE SEE PROFILE

All content following this page was uploaded by Dr M Sakthivanitha on 22 September 2024.

The user has requested enhancement of the downloaded file.

 Artificial Intelligence based System to Prevent Animal
Accident’s in the Railway Tracks

Balaji Kannan R.Bagavathi Lakshmi M.Sakthivanitha R.Maruthi

Assistant Professor, Associate Professor, Assistant Professor, Associate Professor,
School of Computing Science, School of Computing Science, School of Computing Science, Department of Computer
Department of Information Department of Information Department of Information Applications
Technology, Vel’s Institute of Technology, Vel’s Institute of Technology, Vel’s Institute of Hindustan Institute of
Science Technology and Science Technology and Science Technology and Technology and
Advanced Studies(VISTAS), Advanced Studies(VISTAS), Advanced Studies(VISTAS), Science
Chennai, Tamil Nadu, India . Chennai, Tamil Nadu, India. Chennai, Tamil Nadu, India Chennai, Tamil Nadu, India
[email protected] [email protected] [email protected] [email protected]

Abstract - The wildlife population has been in danger by prevention efforts. In Feb 2024, Indian railways introduced an
animal’s mishaps on railroad tracks. According to government AI based system called “AI-based Gajraj” to prevent the death
statistical data, train collisions cause an average of 20 elephant of elephants. The animals have to be saved by implementing a
deaths every year all over the nation. At the same time railways suitable technology incorporating Artificial Intelligence and
departments are taking necessary steps to identify the cause of sensors. The earlier solution uses only sensors to detect the
this and the prevention methods. Despite this, lots of accidents
are happening in the railway tracks majority near the forest
animals that cross or near the railway tracks. In the AI-based
areas. This problem can be addressed by a well defined technology, optical fiber sensors are employed to recognize
technological mechanism needed to alert the animals to be away and track the movements of wild animals at specific areas,
from the railway tracks to reduce the wildlife deaths. In this notifying gatekeepers, control offices, station masters, and
study, Machine Learning(ML) frameworks to detect the animals loco pilots of their movements. The structure of the remaining
that are coming near the railway tracks and provide appropriate part is as follows. The pertinent research on AI-based methods
warning or signals to the animals as well as to the loco pilot. for detecting the wild animals in the railways tracks is
included in Section-2. In Section-3 methodologies used in this
Keywords –Artificial Intelligence, Machine Learning, Animal study are provided and Section-4 discusses the finding and
Deaths , Railways, Logistic Regression, Random Forests, Decision suggestions. Section-5 concludes with a presentation of the
Trees upcoming work and conclusion
I. INTRODUCTION I. RELATED WORKS
The railway accidents are growing daily and it is a Conservation of Animals is very important, and it is to
threat particularly to the wildlife. Almost 115,000 km of track, be dealt with by technology in many ways to make this
or nearly 1.5 times the circumference of the earth, makeup happen. Due to the fact that rails are a commonly preferred
Rail transport in India is among the biggest networks in the method of transportation in Asia pacific countries, the
world. Nearly 20 of India's 101 elephant corridors are known railways are installed throughout forest areas, disrupting the
to have railroad tracks crossing through them, and as natural habitats of the surrounding wildlife. Larger animals
additional tracks are added each year, the issue will only frequently meet their demise when struck by trains. In India,
worsen [1]. Trains are thought to be the second biggest cause these kinds of disasters have become rare, particularly in the
of elephants dying prematurely. Elephant-train crashes pose an southern regions of the country along its lush borders. The
increasing risk to both passengers and elephants as a result of problem is addressed by a computer skill approach that locates
faster trains and increased demand for connectivity. In order to nearby animals using implanted recording devices. It is
address train-elephant collisions, communication with the state advised, tested, and refined to utilize an inception model for
Forest Departments and Indian Railways is necessary [2]. this kind of animal recognition and identification in photos
and videos. These incredibly accurate and accurate
Artificial intelligence (AI) is a constantly developing frameworks have the capacity to warn the trains, potentially
technology with the potential to transform the railroad sector. protecting a precious life. [3]. Deep learning (DL) techniques
AI could be employed to avoid rail accidents involving have proven to be quite effective in identifying the images as
elephants in forest regions, notably those located on either animals recognized on track or not. Convolution neural
mountainous terrain in Tamil Nadu. The primary benefit of networks serve as the foundation for the model that is covered
adopting data-enabled AI is that in the future, the technology in the study discussed in [4] and normalization and pooling
can be leveraged to better comprehend elephant behaviors. layers were employed with the CNN layers to train the model
Given that patterns of behavior are tracked over months, with little loss. Photographs of buffaloes, rhinos, and elephants
seasons, and years, this might enhance the effectiveness of in the wild are included in the dataset. Since public datasets of
railroad tracks and animals are not readily available, the an autonomous onboard system for high-speed train accident,
dataset contains fewer photos. The accuracy of the model is derailment, and fire detection. To find any railroad mishaps or
96% compared to other models using transfer learning. The derailments, a variety of Internet of Things sensors are
method proposed in [5] provides animals that may issue alerts employed, including accelerometers, smoke and flame
that are activated by the trains to reduce the number of wildlife detectors, and more. The MQ2 gas sensor is used to detect
deaths on railroads, especially in locations with a high danger smoke and dangerous gases. Proteus defines the fundamental
of strikes. Similar to the alerts people receive at railroad operation of the accelerometer as simulating the detection of a
crossings, the device blasts a bell and displays lights about 20 collision or derailment by utilizing a potentiometer(3
seconds before the train arrives at the deployment site. variables) for the three axes values. Accurate location tracking
According to learning theory, animals will recognize warning is made possible by the cooperation of an RFID reader
signals when the train is about to arrive if they occur mounted on the train and an RFID tag fixed to the railroad
frequently before it does. Testing is done on two different track.
designs for warning systems: one that wirelessly transmits
data from passing trains to warning devices farther down the YOLOv5 and SENet attention layers were used in the
track, and another that combines warning signals with train building of an effective real-time wild animal detection
system. With the ability to generate auto-alarms, the model
detection in a single device. An analytical model to ascertain
the angular position and features of a live video monitoring uses fine-grain properties of the item to improve detection
system related to length that is intended to be put beside rail accuracy. This allows recognition of a large variety of animals
tracks or roads that cross open, moderately thick, or extremely caught with varying illumination, distance, and background
thick forest zones. In order to prevent accidents caused by fluctuations [12]. A fast method for detecting foreign objects
collisions between animals and trains or vehicles, the in railways is presented in [13]. It is based on knowledge
suggested system has the ability to provide round-the-clock, distillation and a cascaded convolution neural network. By
live monitoring of the rail tracks or roadways and their using a lightweight classification network, the object detection
environs [6]. network can be used less frequently, increasing the total of the
foreign objects. Second, the thin network built has satisfactory
In [7], a multi-sensor environment can identify barriers mage categorization efficacy, since it uses the Overhaul
and calculate their distances in various lighting scenarios for knowledge distillation approach to train it under the
railroad scenes are presented. The system was created as part supervision of a larger network. Lastly, the unfamiliar image
of the H-2020 Shift2-Rail scheme called SMART (Smart- categorized by the first stage network is detected by the
Automation-of-Rail-Transport), which aims to decrease the YOLOv3 object detection network. The work in [14]
number of crashes by recognizing impediments on the rail creatively suggests an early warning system that combines
lines ahead of a moving train. Several deep learning-based event detection with knowledge of the seriousness of un-
models have been carried out in the proposed work to identify familiar objects. In particular, the developed approach can
elephants in images and videos. For elephant detection, a remove the track area using an enhanced railway identification
number of neural networks frameworks and transfer learning appraoch and recognize objects using the cutting-edge
frameworks(ResNet-50, Inception-version-3 and Mobile-Net) YOLOv5 model. The level of caution is established based on
are tested and optimized. Two available datasets, namely the location and seriousness of the obstacles and the hazardous
ELPephant and RailSem19, are used to create a synthetic items are identified.
dataset consisting of around 4200 photos, which is used to
evaluate each model. The suggested CNN has an accuracy of The work in [15] proposes a unique FOD-YOLO net
99.5 and inception network are exceptional at spotting based on DL method to detect items in the railway tracks and
elephants and preventing HEC with higher accuracy of 99.9%. fastener failures. The head and the backbone are the two
For the purpose of preserving other animals in comparable fundamental parts of the deep learning-based YOLOv8. An
circumstances, the same model can be trained [8]. In order to enhanced version of the CSPDarknet53 network is used by
notify the train driver in the event that an elephant crosses the YOLOv8 to identify items on railroad tracks. To detect any
railroad track that the train is approaching, real-time video defects in the track fastening, YOLOv8's brain is made up of
EfficientNet, which has many convolutional layers with
data is analyzed in [ 9]. Elephants are detected from real-time
video data using an adaptive boosting-based Machine squeeze and excitation blocks. These layers are responsible for
Learning (ML) algorithm and the HAAR feature extraction identifying bounding boxes, class probabilities, and objectness
technique. The experimental results suggest that technique has scores for objects and track flaws that are organized with fully
an average precision of more than 96% when it comes to connected layers. Fuzzy logic is used to send the alarm
identifying elephants from real-time video data. message to the loco pilot to avoid mishaps based on data from
the YOLO network. The study in [16] presents a DL and IoT
An object identification technique employing a unique enabled early warning system that runs on solar power and can
feature fusion CNN is used to identify various things in identify stray animals and notify oncoming automobiles of
railroads. For the detecting module that comes next, the their presence. The information about the accident-prone zone
previous module creates the first anchor boxes. The fusion of is delivered by the IoT cloud platform and dashboard. With
features in the object-identification module is used to enhance 91% accuracy, the current system's trained model for detecting
the semantic data for object identification using initial anchor stray animals could predict 86% to 90% of true positive
boxes as input. This enhances the effectiveness of detection, outcomes from real-world camera feeds. The system's 433
particularly for small objects.[10]. Research in [11] suggests MHz RF transmitter can convey alerts up to 110 meters..
 III. METHODOLOGY regression is discriminatory, which means that the algorithm is
genuinely developing the ability to tell the classes apart from
The three supervised ML approaches viz, Logistic one another. Compared to most cats, dogs wear collars, which
regression, Decision trees and Random Forest are used for the is one distinguishing characteristic.
detection process. Logistic regression strategy predicts a
categorical dependent variable from a given collection of
independent factors. A decision tree is a non-parametric
procedure used for tasks such as regression and classification.
Random forest integrates many decision trees to generate a
single, more precise forecast or outcome. Supervised learning
enables users to train classification algorithms to accurately
recognize and differentiate between multiple categories in
training data. This allows specific decision-making bounds,
resulting in precise classifiers.
The objects are any wildlife animals that are detected by
the IoT camera sensors that are fixed in the high striking zones
of the railway tracks. The Sensed image data is then captured
by the local receivers and transferred to the base station
Figure 3.2 AI –Based Technology for classification of animals
through satellites. The image data from the base station is then
received by the corresponding server and identifies the nature The Figure 3.2 shows how the AI based technology is used to
of the object. Then it automatically generates warning signals recognize the captured image data.
to the loco pilot in the train in that route and also to the
roadways near the railway tracks from where the signals B.Decision Trees (DT)
arrived. The signals are capable of stimulating the sirens in Among these approaches, decision trees have found
the roads as well as in the tracks. The siren in the railway widespread application in the classification of aerial and
tracks enables the wild animals to move out of the track only information-based images. It is a straightforward and
if they respond. Otherwise the roaming forest department adaptable computation method as well as an efficient and
personnel or public can take action if the siren goes on the helpful categorization tool for the remote sensing community.
roadways. The Figure in 3.1 shows the overall flow of data Decision trees divide the dataset based on decisions made at
from the various objects to the other one in order to identify each stage. As it is a top-down traversal, the most information
the animals in the rail tracks. should be provided by each split. The main benefit of this
approach is that it can be used to partition large datasets into
manageable chunks when the number of features is high and
the dataset is large. This allows for effective and ideal
classification.
C. Random Forest (RF)
The Random-Forest(RF) algorithm is a prevalent variant of
bagged decision-Trees (DT) that belongs to a family of
ensemble machine learning techniques. The predictions from
several decision trees, each trained on a bootstrap sample of
the training dataset produced by sampling the dataset with
replacement, are combined in bagged decision trees to
mitigate this vulnerability. This method's drawback is that
every tree is trained using the same greedy strategy, and since
some samples may be chosen multiple times during training,
it's likely that the trees have similar or identical split points,
leading to correlated trees. The Random Forest approach uses
a random subset of the training data, produced by randomly
Figure 3.1 Flow of data from the sensors to the server sampling the dataset without replacement, to train each tree in
an attempt to reduce the correlation.
A. Logistic Regression(LR)
This means that each tree must differ from the others
A supervised ML technique that is highly utilized is because the greedy method can only use a certain portion of
Logistic Regression(LR) and it is used for binary class the data to generate the split points that comprise each tree.
categorization problems, in which the outcomes are fixed Every tree in the forest generates a prediction output in the
based on the presence of a specific category of data. In event of a classification problem; the output that most of the
general, a sigmoidal function is used by LR to act on values trees have produced is the final class label. The average of the
between 0 and 1. An automobile or a boat, for example, can be outputs generated by each tree is the ultimate result in a
classified using logistic regression, or it can be used to regression.
categorize more than one class. Additionally, logistic
 IV.FINDINGS AND DISCUSSIONS [2] Nagesh, Tejaswini, Nitin Sekar, and Praveen Nahar. "Using design
methods to identify opportunities to prevent train-elephant collisions in
The experiments were conducted using OpenCV library India." (2020).
to analyse the performance of LR, DT and RF algorithms. [3] A. Ranjith, S. P. Vijayaragavan, N. V and N. Muthukumaran, "An IoT
This experiment uses the dataset from the Kaggle named as based Monitoring System to Detect Animal in the Railway Track using
Animal Image dataset”. It contains about 90 different animals Deep Learning Neural Network," 2022 3rd International Conference on
Electronics and Sustainable Communication Systems (ICESC),
viz: cats, dogs, elephants, horses & lions etc,. The accuracy of Coimbatore, India, 2022, pp. 1246-1253, doi:
each of the algorithms used in this study has been found and is 10.1109/ICESC54411.2022.9885303
tabulated in the Table-1 [4] Khatri, Sandeep, and Jasraj Meena. "Animal Accident Prevention on
Railway Tracks Using Convolution Neural Network." Proceedings of
Table-1 Performance Metric-Detection Accuracy the Third International Conference on Information Management and
Machine Intelligence: ICIMMI 2021. Singapore: Springer Nature
Algorithms/Methods Detection Accuracy Singapore, 2022.
Logistic Regression(LR) 95.6% [5] Backs, J. A. J., J. A. Nychka, and CC St Clair. "Warning systems
triggered by trains could reduce collisions with wildlife." Ecological
Decision Trees(DT) 96.5% Engineering 106 (2017): 563-569.
[6] Acharyya, Aritra, et al. "Analytical modelling of video surveillance
Random Forest(RF) 98.4% system along forest railway tracks." IETE Journal of Education 61.2
(2020): 55-63.
[7] Ristić-Durrant, Danijela, et al. "SMART on-board multi-sensor obstacle
Random Forests tend to outperform single Decision Trees detection system for improvement of rail transport safety." Proceedings
of the Institution of Mechanical Engineers, Part F: Journal of Rail and
for several reasons. I It is an ensemble learning technique that Rapid Transit 236.6 (2022): 623-636.
blends several Decision Trees together. With this ensemble [8] Gupta, Surbhi, et al. "Deep vision-based surveillance system to prevent
approach, the likelihood of overfitting is decreased and a more train–elephant collisions." Soft Computing (2022): 1-14.
reliable and accurate model is produced. Additionally, DT has [9] Dutta, Sayan, et al. "Elephant–railway conflict minimisation using real-
less precise and basic visual forecasts than the RF method, time video data and machine learning." Journal of Reliable Intelligent
which has a complex visualization and accurate predictions. Environments 7.4 (2021)
However, decision trees have low computational [10] Ye, Tao, et al. "Railway traffic object detection using differential feature
requirements, which shorten implementation times and result fusion convolution neural network." IEEE Transactions on Intelligent
Transportation Systems 22.3 (2020): 1375-1387.
in low accuracy.
[11] Guha, Priyadarshi, Sukanya Kool, and Abhijit Paul. "An automatic
In specific use cases, such as managing complex and non- accident detection system for trains using onboard monitoring system."
linear relationships, decision trees are frequently more International Journal of Wireless and Mobile Computing 25.2 (2023):
171-182.
appropriate than logistic regression. Decision trees may handle
characteristics that are numerical or categorical, and they can [12] Bhagabati, Bijuphukan, Kandarpa Kumar Sarma, and Kanak Chandra
Bora. "An automated approach for human-animal conflict minimisation
also manage non-linear correlations between features and the in Assam and protection of wildlife around the Kaziranga National Park
target variable. Moreover the hybrid approaches were used to using YOLO and SENet Attention Framework." Ecological Informatics
enhance the detection accuracy of the algorithms and reducing 79 (2024): 102398.
the number of features in the dataset reduces the complexity of [13] Meng, Siming, Weixun Chen, and Yuelong Jiang. "A Novel Fast
the algorithm. Railway Foreign Objects Intrusion Detection Method Based on
Cascaded Convolution Neural Network and Knowledge Distillation."
IV. CONCLUSION (2024).
[14] Zhang, Zhipeng, et al. "Railway obstacle intrusion warning mechanism
A framework for the identification of wild animals in integrating YOLO-based detection and risk assessment." Journal of
railway tracks has been studied and suggested. The flow of Industrial Information Integration 38 (2024): 100571
data to the different entities and the warning systems provides [15] Brintha, K., and S. Joseph Jawhar. "FOD-YOLO NET: Fasteners fault
a more sophisticated way to detect and avoid animal deaths to and object detection in railway tracks using deep yolo network 1."
some extent. An AI based technology is incorporated to Journal of Intelligent & Fuzzy Systems Preprint: 1-15.
classify or detect the animals using image classifier algorithms [16] Sutar, Vinayak, and Kshama Kulhalli. "An Effective Stray Animal
like Logistic regression, Decision trees and Random forest. It Detection and Crash Prevention System using Deep Learning and
Internet of Things." Grenze International Journal of Engineering &
is evident that the Random Forest approach outperforms LR Technology (GIJET) 10.1 (2024).
and DT methods used in the study. Moreover, the study would
assess the project by examining the moment at which the
system generates a notification and the train comes to a halt,
the number of such warnings generated, whether any false
alarms were issued, and so on. It will take some time to assess
the study and an approach has to be developed with less
assessment and deployment time.
V. REFERENCES
[1] Krishna Chythanya, N., K. Madhavi, and G. Ramesh. "A Machine
learning enabled IoT device to combat elephant mortality on railway
tracks." Innovative Data Communication Technologies and Application:
Proceedings of ICIDCA 2020. Springer Singapore, 2021.

View publication stats