Public Opinion Analysis of Road Accidents&Review of Conventional and Combinational Driver Safety Methodologies with SelfLearning Algorithms

Professor, Department of Computer Applications, Kongu Engineering College, Perundurai, Erode, India. Associate Professor, Department of Computer Applications, Kongu Engineering College, Perundurai, Erode, India. sivabalaselvamani@gmail.com, Associate Professor, Department of Computer Applications, Kongu Engineering College, Perundurai, Erode, India. JuniorResearch Fellow, Department of Computer Applications, Kongu Engineering College, Perundurai, Erode, India.


Public Opinion Analysiswith history ofroad accidents
There are several researchers working on crash prediction to link real-time crash chance with traffic and roadways. Based on the motivation of reducing road accidental death that occurs in Indian roads andas the revision on technologies on VANETs and Machine Learning grows it is strongly believed that applying those will resolve the problem but due to huge road users and road structures it is not that much easy thing to implement in Indian roads. In order to identify the exact reason for accidental deaths, data was collected from 1004 road users out of which 382 female users and 622 female users. We have also collected many repository data from police department and Health & welfare department.

fig. 2.1.
The total number of accidents in Tamil Nadu urban and rural roads, which are recorded with SCRB, Chennai (responsible for data in police department) is shown in figure 1 which has the highest of 71,432 accidents in the year of 2016 and upon that the count is in and around 75 thousand approximately which is found once discussing with the police chief.  1994 1994-1995 1995-1996 1996-1997 1997-1998 1998-1999 1999-

Change in % of fatalities for every year
It is observed that, there is a steady increase in the growth of fatalities due to accidents of 3.2% as an average of every year. 2003 has around 7% lowest death rate recorded compared to 2002 has been seen as good scenario and it was not maintained because in the next consecutive year of 2016 it has found that 12% of increase compared to 2015. From this implications it can be concluded that the death rate due to accidents will vary accordingly year by year based on several parameters.
We classify accidents into four types based on injuries occurred in accidents, Fatalities, Severe injuries,Minor injuries and non-injuries.

fig 2.4.
As compared to any other three factors, minor injuries are higher which also denotes that the accidents are still happening even after this much researches and implementations in road safety and vehicular technology and infact vehicle accidental death have multiplied by three times from 39,675 in 2003 to 90,729 in 2017. Tamil Nadu recorded second highest road accidents in India in 2017. A questionnaire is formed to collect the public opinion about road accidents which was collected from 1004 persons is given with the support values in

46.
There is no particular accident detection system in our society 70.7

47.
A alternative traffic control system designed especially for ambulance is needed in India 69.3

48.
Road safety awareness camp is needed at regular intervals 67.9 Table 2.1 Support value is initially collected from public based on individual's score of that statement for 5 and then converted to 100.

Theory behind prevention and detection of accidents
The relative need of efficient driver alert system and automatic accident detection is an action of precaution and postulation of the accidents in road safety. To understandthe realistic relationship between prevention (Precaution) and detection (Postulation) of accidents the following theory has to be followed. D ∝ 1 P ….(1) P is Prevention of Accidents andD is Detection of Accidents.

Assistance from various devices for Accident Detection
The devices involved in Advanced Driver Assistance System (ADAS) are Monocular camera, Stereo camera, Fish eye camera, Charge couple device camera, RADAR, LIDAR, MEMS, NEMS, Gyroscope, IR camera, EEG, GPS, GSM,Multimodal sensor, Vibration sensor, Speed sensor, Alcohol sensor, Proximity sensor, Ultrasonic sensor,Photonic mixer devices and so on. All these devices provide safety to driver but stillaccident persists due to failure of system or driver carelessness.Several accident detection techniques are available anditis activated when the driver is not able to get alert from ADAS and meet with an accident. When accident occurs,transforming theinformation from that accident spot to the rescue center for further action is the biggest research challenge in current days.

fig.3. 1a
fig. 3.1b It was introduced by Dooley et al [1], On the rear side of the car, a fish eye camera is positioned to identify the target vehicle (TV) entering the blind zone using calibration technique and monitors the vehicle with the aid of the angular velocity and speed of that vehicle after entering the blind zone, as seen in Fig 1b.The data received from camera is classified using Adaboost classifying technique to identify the TV from 10-40m before vehicle enters into blind zone. Identification of all the zones was a challenging research in which Google cars has produced high potential output using mounted-LIDAR units.
The vehicle safety mechanism with the help of Suspension and Control System (SCS) is an approach of Automatic Cyber Physical Systems (ACPS) that provides knowledge to the vehicle in all zones and is proposed by Naufal et al [75].It uses four proximity sensors to detect moving vehicle from front, back and sideways [75]. A 2 CPS Engine is an embedded device fixed in the vehicle that receives data from sensors, other vehicle and from information centre. It takes all computational work and providesIncident Zone Warning (IZW), a safety mechanism used to provide alerts based on the features extracted from the proximity sensors. A few other computer vision methods that are used in road safety applications in Table 2. The speed factor plays an important role in the study of the causes of accidental deaths; fifty drivers were recruited by Charlton and Starkey to analyze the driver speed choice and it was found that the mean speed is normal among all drivers in varying road situations [93]. Prototype willingness model is developed by Preece et al. [92] to identify fast driving and willingness in text driving is conducted in IBM SPSS Version 24 form the data received from 183 licensed drivers of 17-25 years old drivers and found that most of the positive attitude young drivers were protagonists. Several other factors that cause accidents are red light running, fog or snow, pedestrian crossing, collision of the animal, driver fatigue, multiple vehicle collision, brake failure, unlicensed driving, high beams in NH and SH, distracted driving (not following traffic rules),speeding in turns and un safety stunts [ [4]. This system has major role in densely populated city where animals like dogs, cows, monkeys and donkeys are large in number. In the CV process, the Histogram of Directed Gradients (HOG) is used to define the image characteristics [4]. It is trained under 2200 images. It solves big issue with less accuracy.Farid et al developed Safety Performance Function that computes crash counts for rural areas in four states [84] and one of the hardest predictive problems is critical night time accidents [99] thatoccurbecause of dangerous curves when driver is unaware of it. These curves are highly severe at worst climatic conditions like fog.

Road Curvature Estimation using Circle Fitting Algorithm
Lee et al [13] introduced a method for estimation of road curvature using W-Band FMCW RADAR and Circle FittingAlgorithm that works in all possible critical conditions.The minimum orthogonal squares to minimize the sum of squares in the algorithm of circle fitting are given by Where f can be defined by geometric and algebraic models. Lee et al uses algebraic model, the popular algebraic equation for Circle Fitting Algorithm is ( 2 + 2 ) + + + = 0 The minimization of the simpler function is ( , , , ) = ∑( + + + ) 2

=1
Where 1/( 2 + 2 − 4 ) = 1 and function F in matrix is given by = ‖ ‖ 2 = ( ) A is parameter vectors and X is (zi,yi,zi).Lee et al uses function F with time varied vehicle dynamics as main input that is computed by Controller Area Network (CAN) and provided as input for radar sensor through CAN bus. The procedure provides 10-17% more accurate prediction compared to map data.

3.3Driver Safety System
Driver Safety System (DSS) fall in two categories, Active and Passive Safety Systems (ASS and PSS). ASS provides Advanced Driver Assistance Systems (ADAS) that provide assistance by automatic collision control, braking and lane keeping. PSS are air bags and seat belts kukkala et al [79]. Camera used in computer vision technique is of two major types one is monocular (have only single lens) and other is stereo (more than 2 lens). Stereo cameras provide 3-D data by combining two or more images captured from various lens and it is mainly used for ASS. There are other devices that are used to get input data for ADAS and Accident Detection System (ADS) like IR cameras, LIDAR, RADAR, Ultrasonic sensors and photonic mixer devices. Each pixel of the frame is represented in the form of matrix data. ADAS processes these data around 60 fps in which each frame contains 3 channels (RGB).
It undergoes various processes, such as segmentation, identification and monitoring of objects, measurement of depth and control of structures. Machine learning algorithms such as LR, Gradient-Decent, SVM, k-NN, ANN and CNN gain from ADAScomputing [79].A driver fatigue recognition model works with featuresextracted from ECG and EEG is proposed by yang et al [2]. Hidden Markov Model (HMM) is used to compute Dynamic Bayesian Network (DBN) at various time stamps is the Collection of Static BayesianNetwork (SBN) that is interconnected, the relationship that interconnects SBN is found using 1 st order HMM whileDBN classifies output to alert the driver. The importance of Electroencephalography in Sleepiness Detection System (ESDS) is given by Balandong et al [80]that act as interface between the human brain and control unit. EEG uses electrodes, these electrodes are fixed at nearest part of human brain to measure EEG signals[80] that extracts features from f8-electrode of EEG then used for SVM and Bayes for supervised learning to provide higher results.
Six key classified features used to define performance include different SDS techniques 1)subjective assessments 2) vehicle-based systems 3) behavior-based driver systems 4) Sleep-Wake mathematical models (MMSW) dynamics 5) human physiological signal-based systems 6) combinations of one or more of these techniques [80]. The conventional technique used to measure the level of sleepiness of the driver and steering wheel movement is the Karolinska Sleepiness Scale (KSS), percentage of eye close (PERCLOS), Bio-Mathematical models are the techniques that plays a major role in the EEG evolving models.

3.4Autonomous Vehicles
Deployment of software and electronic innovations in automotive industryfor connected vehicles is surveyed by siegel et al [78]. Applications of RADAR and LIDAR towards perception to find angular velocity of vehicle are used in autonomous vehicles which in turn increases the production of vehiclewith the aim to satisfy and safeguard the customers. RADAR Sensory methods are emerging with luxurious monster vehicle producer like Tesla, Google-Autonomous Vehicle, BMW and Benz where the connectivity of autonomous vehicles of inter and intra vehicle is important in designing of any autonomous vehicles. Sharing of sensor and actuator data is most needed in intra vehicle connectivity, whereas inter vehicle networks share the critical orderly information to another vehicle through VANETs. IEEE 802.11 P is the unique communication protocol used in V2X applications.

Enhancement of V2X Communication using Multi-RAT
Vehicle to everything communication provides high reliability in processing important information in 5G Cellular networks. Single transmission techniques like LTE-Uu or PC5 will not help vehicle to everything communication, so a Multi radio access technique is introduced by J. Lianghai et al [14]. For Multi-RAT, a hybrid uplink mechanism is designed to boost reliability, it is designed to accept both LTE-Uu and PC5 packets in such a manner that OBU allows to relay the signal from the vehicle to the uplink. Accident detection is computed using On-Board Unit (OBU) which passes accidental information using VANET to reach the concern [16]. Ki and Lee developed a model to predict accident at intersections using image captured charge couple device camera and feature extracted from moving vehicle is passed to ARRS model [7]. ARRS model analysis and evaluation of captured image is carried out with standard DVR to record and send information from AMP to TMC. Three major steps taken in this algorithm are vehicle extraction, feature extraction and detection of accidents to compute acceleration, position, area detection and compared with threshold to formulate the incident. In a recent paper Yaun and Abdel-Aty proposed a method to identify the accidents at Intersections and Intersection entrances [86]. Some of radar, lidar and sensory methods available are listed in Table 2.

3.6Traffic Detection (TD)
The road congestion detectionto control the level of congestion is proposedto overcomemany traffic prediction techniques that failed to support the flow of traffic is provided byKalamaras et al [77]. As a parametric model, autoregressive integrated moving average (ARIMA) is used and k-NN, SVR is used as a non-parametric model to measure the efficiency of both models for traffic prediction. The accuracy of prediction is carried out by normalized root means square error. An IoT cloud system to provide the traffic analysis is introduced by celesti et al [76]. Fixed sensors sometimes fail to send data through 4G network. This system sends data directly from vehicle to cloud with the help of IoT. Cloud system receives data and processes using #openGTS to provide information about traffic to the passers in the same route with detailed information through GPS. A visual analytics tool is developed and maintained in the cloud server.

Hybrid approach (CNN inVANETS)
Convolution neural network has large number of parametersand this serves as an efficient approach in any kind of recent decision making system. Deep learning technique of Convolution Neural Network is used in computer vision and In the NLP method, the Recurrent Neural Network is used. An powerful method of crash prevention would be the combination of the CNN in Vehicular ad-hoc network to make crucial decisions in accident circumstances.
In the recent days most of the computer vision techniques follow CNN. Pixels of an image is represented as value matrix, based on the intensity of a particular pixel, the resulted huge matrix is given as input to convolution layerin Image Recognition (IR) problem. System uses 3D volumes of neurons (Width*Height *Depth) for further computation,For the identification of car license plates, a CNN-based MD-YOLO system has been implemented. The prediction of rotation angle and intersection over union is used to treat real-time conditions and to achieve better results [3]. Deep learning method is implemented in the framework for lane detection using multi model sensors that is capable of receiving the input in various situations [10].The principle of computational experiments and simultaneous execution (ACP) is used to develop this method to understand, implement and test the lane detection mechanism in artificial society [10]. In CNN, input layer, convolution layer, induction layer, pooling layer, and output (fully connected) layer, there are usually five layers. Others are opaque layers, except for input and output layers.
The output obtained from the CNN chooses to trigger the contact channel. The simplified crash details would be transmitted from the car to everything through the contact channel (V2X). In Vehicle to Everything communication there are several issues that are still not properly addressed in VANET. They are congestion of data from Vehicle to Vehicle (V2V), security of the vehicle data when it is connected to everything which will make the VANET to perform poor. To solve this, all the vehicles should need to get connected only through the base station; warning messages are preferable from the base station to avoid accidents. Activation of vehicle to vehicle communicationwithout the support of base station should only beafter the accident to ensure high level security in VANETs. Deep learning and hybridization techniques that are applied in the field of Transportation Systems andDriver Technology will be the enhancement of ITS.

Results and Discussion
As a part of opinion analysis in this paper, it is found that the support value received for Q1 is 84% and after that there was the steady decrease in the support value for each and every question. This also shows that people in Tamil Nadu state are little careless about road accidents and thus recorded the second highest road accident fatalities in 2017 among Indian states.
In DSS,Driver Drowsiness Detection (DDD) model developed by Li and Chung [21] used to detect drowsiness produces 96.15% accuracy that is higher than any other Drowsiness detection model and acceptance of driver alert system by driver is assessed by Rehman et al and produced 85% variation in results [90] which gives very low results when compared to Li and Chung[21]. The study of Pylkkonen et al [88] with 53 truck drivers in the intervention to provide the sleepiness countermeasures is failed and proves that driver education is not the sufficient measure for driver sleepiness.
These two results infer us that neither driver education nor driver alert system is sufficient to protect drivers. Accidents are exceptional and even though there are high end auto pilot cars like Tesla s model and Benz F 015, there is still a huge gap in prevention of accidents and it has been proved once Tesla S model failed to protect its driver who met with an accident in auto pilot mode. Thus in order to prevent accident an exceptional hybrid model is suggested in this paper.

Conclusion
This paper provides a detailed study to clarify the significance of the accident warning mechanism in Tamil Nadu. Various devices, techniques and methodologies that are used in ITS is detailed in this review with the most 4 influencing algorithms in the ITS which is more popular in various developed countries and several experimental resultsDifferent papers demonstrate that, using the new algorithms and processors such as GPGPUs, the standard computer vision approach can produce more accurate and precise performance.Based on the result of this study the hybrid approach of CNN in VANETs considering security could be the best research suggestion to avoid accidents.