Volume 5, Issue 5, May – 2020 International Journal of Innovative Science and Research Technology

ISSN No:-2456-2165

Detection of Helmetless Riders Using Faster R-CNN

Meenu R1, Sinta Raju 1, Smrithi P Paul 1, Swathy Sajeev 1, Alphonsa Johny 2
Department of Computer Science and Engineering
St Joseph’s College of Engineering and Technology Palai, Kerala
1
B.Tech Student
2
Assistant Professor

Abstract:- In India, more than 17 deaths and 55 road registration number is recognized from the vehicle using
accidents happen each and every hour and mostly the open-ALPR and an alert is sent to the nearby police station.
two-wheelers add the account to an increase in road
accidents belonging to the age group of 18-45 yrs and Here we mainly categorized into four areas, they are
the causes regarding the same might be due to the  Motorcycle Detection
helmet-less riding, triple riding and speeding. To  Head Detection
overcome this issue we have come up with an idea to  Helmet Detection
detect the riders without a helmet, those having a triple  Number plate detection
ride and to inform the cops when an accident occurs,
this project mainly lends a helping hand to the cops. II. BACKGROUND
The cause of accidents by the two-wheelers is mainly
because of not wearing helmets. Traffic control officers Since bikes are reasonable and a day by day method
and doctors often say that the death could have been of transport, there has been a quick increment in
avoided if the rider was wearing a helmet. The riders bike mishaps because of the way that the majority of the
wear helmets when they see a police vehicle or a motor motorcyclists don't wear a helmet which makes it an ever-
vehicle checking ahead of them. The number of deaths present risk each day to go by a bike. Over the most recent
or accidents caused by the shock of a sudden reaction few years alone most of the deaths in accidents are due to
when a helmetless rider sees a cop vehicle is damage in the head resulting in trauma to the skull or mind.
remarkable. Another cause of accidents is taking more In light of this, wearing a helmet is obligatory according to
passengers than the manufacturing company of the two- traffic rules, violation of which pulls in heavy fines. Inspite,
wheeler. We propose a framework for real-time an enormous number of motorcyclists don't comply with
detection of traffic rule violators who ride bikes without the standard. The police officer attempted to control this
using a helmet. This methodology initially identifies issue physically, however it is insufficient for the real
bicycle riders from surveillance video background circumstance. The requirement for security measures is an
subtraction and object detection. At that point it decides unquestionable requirement to decrease the number of
if the bicycle rider is using a head protector or not using deaths in road accidents, and use of helmets is a significant
Image Processing. In the event that the rider is riding factor regarding safety.
helmetless, at that point the number plate of the bicycle
is noted and the traffic safety rule violation report with A study that was conducted by the United Nations in
the licence plate number and the picture is sent to the 2015 estimated that the chances of surviving an accident
particular cop to make the move required. rose by 42 percent on wearing a helmet. Even though
helmets are for the safety of the riders, most of them avoid
Keywords:- Image Processing Algorithms, Object it due to reasons like “it spoils my hairstyle”, “it feels
Detection algorithms, YOLO Dataset. uncomfortable”,” good helmets are costly” or “it obstructs
my peripheral vision”. These reasons are not comparable to
I. INTRODUCTION losing a life. The existing system for checking whether a
rider is wearing a helmet or not is a checkpoint by police or
In India, the number of accidents occurring each day other personnel to manually check each rider. In this
is increasing rapidly. The two-wheelers account 25 percent system, there is an impossibility of riders evading
of road crash deaths because of ignoring safety measures checkpoints.
like wearing helmets while driving. More than two drivers
are travelling in a two-wheeler is also a major reason. So Thus the importance of automatic systems in traffic
here, we propose a framework for real-time detection of control has been increased in recent years. Presently, all
traffic rule violators who ride bikes without using a helmet. major urban areas already deployed huge video
The road CCTV footage is used to detect whether a rider is reconnaissance systems to keep a vigil on a wide
wearing a helmet or not, using Deep Learning and Image assortment of dangers. In this way utilizing such a already
Processing technology. The algorithm used here is faster R- existing system will be a cost-efficient arrangement,
CNN, because it increases the detection rate of however, these frameworks include an enormous number of
motorcycles, compared to other deep networks such as people whose performance is not significant for long
CNN, fast R-CNN and YOLO. The violator’s vehicle’s periods of time.Recent studies have shown that human
surveillance proves ineffective, as the span of checking of

IJISRT20MAY841 www.ijisrt.com 1616

Volume 5, Issue 5, May – 2020 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165
recordings expands, the blunders made by people likewise CNN network to detect the bounding box area of
increases. We aim to improve the usage of a traffic stream motorcycle and rider and classify if the biker is wearing a
framework, others are to lessen the expense of human work helmet or not. The results of the experiment were good as
and abate the reasons for a mishap. The ideal solution is to the Deep Learning and CNN techniques were good
develop an electronic detection system that can be algorithms which could be applied for the problem of
automated to recognize this kind of problem without human image detection.
cost.
Yogameena’s[12] paper of helmet wear analysis using
III. RELATED WORKS deep learning had the framework which used the
performance metric and means average precision on the
Based on the papers we did for research, a short CCTV footage. The framework consisted of foreground
analysis is done: In the paper done by Vishnu[7], a system segmentation with the help ofGMM, motorcycle detection
for the automatic detection of the motorcyclists driving and detection of motorcyclists with or without helmet using
without helmets in the CCTV footage. Versatile faster RCNN
background subtraction was utilised for the video edges to
get the moving object.CNN was also used for identifying In Yuanlong Yu’s[13] paper, they developed a study
the motorbike rides in the moving object. They were able to for a traffic accident detection method which includes a
identify 92.87 percent of the violator with a low false alarm self-tuning iterative hard thresholding algorithm for
rate of 0.5 percent on an average and with these lines learning sparse Spatio-temporal features and a weighted
showcased the proficiency of their proposed methodology. extreme learning Machine for detection. The drawback seen
here was that false detection occurred when the collision
Dharma Raj[9] also showed a related work of the happened in different depth.
developments of a system using image processing and deep
convolutional neural networks(CNN) to identify the In Rohit’s[14] paper, a similar method of helmet
motorcyclists who violate the helmet law. Their system detection was developed using deep learning, they used the
consisted of motorcycle detection, helmet v/s no helmet Caffe Model for the detection and extraction which had an
detection and motorcycle license plate recognition. CNN accuracy score of approximately 86 and another method
models were built for the classifier. The errors they faced used for image classification was the Inception V3 model
were when the rider wore a hat it was identified as a helmet which has an accuracy score of 74.
classifier. This issue could be resolved by increasing the
training data with hats. IV. PROPOSED METHODOLOGY

Baolin Bai’s[8] paper, the system was developed with The system is divided into four modules:
the Deformable Parts Model(DPM) which has a high
detection accuracy in the image detection algorithm. The A. Motorcycle Detection
DPM is both efficient and accurate for object detection. B. Head Detection
The algorithm of MPCA and fast level locating was used C. Helmet Detection
for the training process. Even though high detection D. Number Plate Detection
precision and detection speed were enhanced, the algorithm
in the paper could only deal with the partial occlusion or
damage with the vehicle occlusion problem and the effect A. Motorcycle Detection
of detection was not accurate under the occlusion of Bike and helmet recognition from a picture has been a
vehicle’s condition or damage seriousness. False detections precarious circumstance in the field of image processing.
happened due to confusion among classes such as between The difficulties confronted were the state of the bike in the
car and bus and in other categories false detection was picture, the acknowledgement of individuals riding on a
often due to the relatively strict bounding box criteria. cruiser or an empty vehicle with no rider, the spot of the
traveler’s head and the distinguishing proof of the helmet at
Nagarjuna’s[10] paper had developed a car accident the head area of the rider. Various steps of image
detection and communication system which would help to processing were applied on the video images, before it
inform the relatives, nearest hospitals and police with the detects the motorcyclist in the frame and also it’s noticed
location of the accident. The system was able to send that the properties derived using various algorithms using
messages to the stored emergency numbers when the car the information present in the image itself provide decrease
collided or topped or tilted by more than 30 Degree. A the detection rate for motorcycle detection. Recently, deep
drawback identified in this system was that the coordinates learning algorithms such as CNN have taken on helmet
of the location were sent instead of the exact location which wear analysis with it is comparably better than such hand-
could be done by an application locating the map’s crafted features. As the faster R-CNN escalates the
location. recognition rate of motorcycles, compared to other deep
networks such as CNN, fast R-CNN and YOLO. We use
In Narong’s[11] paper, a method using deep learning faster R-CNN. This research and relative study for selecting
called Single Shot Multibox Detection was developed into an admissible algorithm, which provides more desirable
the helmet detection problem. In this method, only a single accuracy in the existing world challenging conditions in

IJISRT20MAY841 www.ijisrt.com 1617

Volume 5, Issue 5, May – 2020 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165
difficult surveillance data and they are the key benefaction depends on the position of the bounding boxes that is
of the proposed work. We use the CCTV footage videos created around the motorbike and rider.The detected
from the desired locations and we divide the video into motorcycles were first manually classified into helmet and
different frames later we detect objects using open-CV non-helmet classes for training the device.For training, we
source python code(based on faster R-CNN) and if the use the upper one-third of each motorcycle images as a
system detects the presence of a motorcycle, it first checks region of interest extracting process,because this is the
whether it has a rider or not. Based on this,the desired part region most likely to contain the helmet or no-helmet
from the frame and further checks are done. informations. This new set of coordinates is used to crop
the current frame.These cropped images from the frames
B. Head Detection that is motorbike person combination.This image is then
From the above frames, we distinguished a cruizer passed onto the image classification model then classifies
(utilizing faster R-CNN). On the off chance if the model the cropped images into a helmet or no helmet classes and
distinguishes the nearness of the bike, a limit box is made saves the image in directory else it is simply
around it. This zone is then withdrawn from the current discarded.Another key contribution of the proposed system
frame and passed to the image classifier for additional includes most of the real-world challenges of the head or
processing(image acquisition, image restoring, linear helmet detection such as low resolution,blur,profile
filtering,etc.) in the later stages. The image classifier will view,occlusion,bald head and persons with different helmet
isolate the test image from the captured frame into one of shape and size.The dataset collection and experimentation
the two goal classes. Faster R-CNN with VGG-16 a pre- on these datasets are other major beneficiation of the
trained model could distinguish humans or objects like proposed work.
person, horses and chair. In the view of camera position
with its angle of depression, it produces human-head D. Number Plate Detection
appearances. With the assistance of Gabor-Wavelets filter, In continuation with the detection of motorcycles, the
conforms its sturdiness and stability against changes in the detection of motorcyclists with and without a helmet is
scale, orientation and dazzle are used for identifying performed similarly. We have experienced several different
highlights representing the facial segments. It steadily CNN models for character classification of motorcycle
identifies the human head under various ecological licence plates. We prepared the framework with various
conditions. pictures of helmetless riders and we assembled the relating
number plate. That is the motorcyclist with the protective
C. Helmet Detection cap isn't contemplated however the frame with the biker
It is inferred from the related work, that the features without head protector is put away in the registry. These
and the CNN that exist now fail to control the real-world pictures are moved to further processing. The entire system
challenges for helmet wear analysis such as helmet has been tested by expelling this module and the rest of the
detection of motorcyclists, so we use faster R-CNN. To groupings, for example, bike detection and recognition of a
find a robust and accurate model for helmet or no helmet motorcyclist with or without a helmet has been actualized.
classification, we begin our project from a simple model Hence, the region related (false positive) to the motorcycle,
and increase the complexity of the model incrementally motorcyclist, head and helmet in background regions and
with input channels of RGB colour. The identification of foreground regions have been detected. As a result, it
people riding a motorcycle or it’s just an empty vehicle proves the significance of the initial foreground
with no riders, the location of the biker head and the segmentation, motorcycle detection, detection of a
detection of a helmet at the head location of the rider was motorcyclist with and without a helmet and Licence Plate
examined. Various steps of image processing needed to recognition of the motorcyclists with and without helmets,
apply on the video frames before it can identify the position delivers a framework for an automated helmet wear
of the motorcycle. In this module the detection model analysis system. The study of related work, dataset and
detects the presence of a motorbike and person with a experimentation on different types of challenging
helmet. The rider overlapping area at a specific point of conditions are the key beneficiation of the proposed helmet
time is detected and a set of coordinates and dynamically it wear analysis framework.

IJISRT20MAY841 www.ijisrt.com 1618

Volume 5, Issue 5, May – 2020 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165

Fig 1:- The framework of the system

V. RESULT

The results obtained are shown as Figures below.

Fig 2:- Detection of Bike

Figure 2 shows the result of detection of the bike using our system. Bikes are highlighted with green colour boxes.

Fig 3:- Bike Rider without helmet detected

IJISRT20MAY841 www.ijisrt.com 1619

Volume 5, Issue 5, May – 2020 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165
The main part of the system, that is the detection of [4]. C.-C. Chiu, M.-Y. Ku, and H.-T. Chen, “Motorcycle
helmetless riders are shown in Figure 3. Here the head with detection and tracking system with occlusion
the helmet is highlighted with blue boxes and that of the segmentation,” in Int. Workshop on Image Analysis
head with no helmet is highlighted using red boxes. for Multimedia Interactive Services, Santorini, June
2007, pp. 32–32.
[5]. J. Chiverton, “Helmet presence classification with
motorcycle detection and tracking,” Intelligent
Transport Systems (IET), vol. 6, no. 3, pp. 259–269,
September 2012
[6]. Z. Chen, T. Ellis, and S. Velastin, “Vehicle detection,
tracking and classification in urban traffic,” in Procs.
of the IEEE Int. Conf. on Intelligent Transportation
Systems (ITS), Anchorage, AK, Sept 2012, pp. 951–
Fig 4:- Cropped Number Plate
956.
[7]. C. Vishnu, Dinesh Singh, C. Krishna Mohan and
After detecting the helmetless riders, their vehicle
Sobhan Babu,”Detection of Motorcyclists without
number plate is cropped (Figure 4) and sent through the
Helmet in Videos using Convolutional Neural
mail to the authorities along with the frame. Network” in 2017 International Joint Conference on
Neural Networks
VI. CONCLUSION [8]. Baolin Bai, Yusong Liu,Chunyang Zhang , “Accident
Elements Detection based on Improved DPM” in 2017
This project aims to decrease the accidents caused due
32nd Youth Academic Annual Conference of Chinese
to not wearing the helmet. It also ensures whether the law is Association of Automation.
violated or not. It stores the image of the violated people [9]. Dharma Raj KC, Aphinya Chairat, Vasan Timtong,
along with a cropped number plate. It is then sent through Matthew N. Dailey, Mongkol Ekpanyapong, “Helmet
the mail to authorities. The algorithm used here is the
Violation Processing Using Deep Learning” in 2018
Faster RCNN which is faster and accurate when compared
International Workshop on Advanced Image
to the other one. The accuracy of the result of this project is
Technology.
92%. It depends on the quality of the CCTV camera. [10]. Nagarjuna R Vatti, PrasannaLakshmi Vatti, Rambabu
Vatti, Chandrashekhar Garde, “Smart Road Accident
ACKNOWLEDGMENT Detection and communication System” in 2018 IEEE
International Conference on Current Trends toward
Behind this successful venture is the blessings and
Converging Technologies.
guidance of many people. First and foremost, we thank God
[11]. Narong Boonsirisumpun, Wichai Puarungroj, and
Almighty, who gave us the inner strength, resource and Phonratichi Wairotchanaphuttha, “Automatic Detector
ability to complete the work successfully, without which all for Bikers with no Helmet using Deep Learning” in
our efforts would have been in vain.
2018 22nd International Computer Science and
Engineering Conference.
We are deeply indebted to Ms. Alphonsa Johny,
[12]. B. Yogameena , K. Menaka, S. Saravana Perumaal,
Asst.Professor, Computer science and Engineering “Deep learning-based helmet wear analysis of a
Department, for giving necessary suggestions as a project motorcycle rider for intelligent surveillance system”
Guide. in IET Intelligent Transport Systems (Volume: 13,
Issue:7 2019)
REFERENCES [13]. Yuanlong Yu1 , Miaoxing Xu1, Jason Gu2, “Vision-
based traffic accident detection using sparse spatio-
[1]. Kunal Dahiya, Dinesh Singh, C. Krishna Mohan, temporal features and weighted extreme learning
“Automatic Detection of Bike-riders without Helmet machine” in IET Intelligent Transport Systems
using Surveillance Videos in Real-time” in 2016
(Volume: 13, Issue:9 2019)
International Joint Conference on Neural Networks.
[14]. Rohith C A, Shilpa A Nair, Parvathi Sanil Nair, Sneha
[2]. A. Adam, E. Rivlin, I. Shimshoni, and D. Reinitz,
Alphonsa, “An Efficient Helmet Detection for MVD
“Robust real-time unusual event detection using
using Deep learning.” in 2019 3rd International
multiple fixed-location monitors,” IEEE Transactions Conference on Trends in Electronics and Informatics.
on Pattern Analysis and Machine Intelligence, vol. 30,
no. 3, pp. 555–560, March 2008
[3]. B. Duan, W. Liu, P. Fu, C. Yang, X. Wen, and H.
Yuan, “Real-time on road vehicle and motorcycle
detection using a single camera,” in Procs. of the
IEEE Int. Conf. on Industrial Technology (ICIT), 10-
13 Feb 2009, pp. 1–6.

IJISRT20MAY841 www.ijisrt.com 1620