Machine learning based IoT system for secure traffic management and accident detection in smart cities

Introduction

Worldwide, there are 1.35 million mortality cases, accounting for $2.2$ percent of the global population, reported each year. Traffic accidents lead to deaths and permanent impairment. Speeding, careless driving, driver exhaustion, stray animals, and inadequate infrastructure are the primary causes of these accidents. Emergency medical services' slow response times exacerbate fatalities and impairments.

The "golden hour," the time after a traumatic injury, is crucial. Immediate surgery and medical care during this period raise the likelihood of survival by one-third.

IoT for Information Technology Management (ITM) is connected to automated vehicles, cooperative transportation systems, and intelligent roadways, improving data transmission, creating diverse connectivity, and enabling low-bandwidth gadgets in high-capacity locations. According to NSO research, India's GDP will decline by $7.7$. ITM systems can resolve traffic management difficulties, improve logistics and passenger transportation by reducing traffic congestion, and raise people’s living standards. ITM initiatives must include smart-city governance, which develops planning techniques for improved laws. A critical tenet of smart-city management systems involves innovative services.

IoT networks use cutting-edge technologies like ITM to connects physical objects to the internet and establishes intelligent networks and wireless transmission connections. Communication among IoT-based vehicles is a novel information exchange model that supports ITM. IoT combines data collection, sensor data processing, and computation to handle and sustain traffic networks. Autonomous transportation employs traffic lights with timers for each phase, as well as electronic detectors. Despite computerized traffic-control sensors, road traffic persists. Intelligent transportation systems can resolve traffic jams and other challenges.

Accident Alert Sound System (AALS)

This study proposes an AALS system to identify car accidents and notify nearby automobiles without requiring modifications to new or EVs. The AALS system uses the event-driven wireless sensor network (EDWSN) protocol and functions independently to communicate among nodes in the event of an incident. The AALS system recognizes an accident using its sensing capabilities and delivers light and audible alerts.

ATM Based on DL and IoT

Intelligent cities leverage AI, DL, and IoT-based systems, like embedded linked devices, IoT sensors, intelligent traffic-control systems, streetlamps, and roadways, to collect and analyze data for optimizing public utilities, infrastructure, and communication across various platforms, including Industry 4.0, intelligent buildings, smart cars, agricultural production, and innovative healthcare systems.

The main contribution of the proposed method is given below:

• The proposed method uses an accident alert sound system (AALS) to detect accidents and alert vehicles in smart cities.

• Identifies current traffic circumstances and identifiable patterns in the flow of traffic, allowing SEE-TREND to forecast impending traffic events and notify relevant parties of their tendency to occur.

• One of its key advantages is the capability of the suggested ATM-ALTREND structure to interact with any adaptable approach without requiring changes to the traditional architecture.

Related Works

Intelligent Transmission-Control System

Bhatia et al. (2022) described an intelligent transmission-control system using cloud viewpoint and ML techniques. The cloud picture API is used to identify the concentration and driving experience. Traffic intersection’s visuals are recorded and preserved in the cloud database. The situation is transferred to the next traffic light. The previous traffic signal, which is now operating, will monitor how the next traffic light performs and then carry out the activity by the circumstances.

Traffic Behavior Anticipation

Kaginalkar et al. (2021) showed that the existing techniques could assist in anticipating traffic behavior, automated traffic-signal administration, parking recognition, and detection of surrounding objects/vehicles, which can boost the safety and effectiveness of ITM. Improving traffic control is still challenging.

IoT-Based Intelligent Traffic-Control Systems

Increased traffic monitoring systems have been shown by the writers of Gatto & Forster (2020), to alter bright metropolitan regions. According to Cao & Wang (2020), automatic traffic detection is the foundation of services and facilities for urban planning. Intelligent connection sensor networks calculate traffic flow, foresee traffic snarls, and adaptable manage traffic flow, raising awareness and equipment use.

Vehicle Congestion Assessments

Schneider, Lugner & Brandmeier (2019) reported that road traffic, utilization, and average densities are extensively used in vehicle congestion assessments. Most of this data was gathered from images and videos captured by machine vision software.

For this specific situation, the authors of Sequeira et al. (2020) presented an IoT-enabled monitoring system to collect, run, and consolidate real-world traffic patterns. The researchers suggested an IoT-enabled control system (Bugeja et al., 2020) to acquire, manage, and aggregate real-world traffic conditions to improve motion range and disseminate traffic information about traffic jams and crashes via the highway signaling system.

Traffic State Usefulness

A framework was used by Javed, Zeadally & Hamida (2019) to investigate the usefulness of the traffic state. The testing results show slight inaccuracy in the forecast of highway occupancy and good vehicle detection and tracking precision.

Deep Automated Learning (DL)-Based System for Accident Detection

The development of deep automated learning (DL)-based system (Tian et al., 2020) for accident detection using visual information. The method uses temporally ordered graphic elements to illustrate traffic crashes. The system design consists of a phase for extracting visual features and identifying transient patterns. During the learning phase, sensory and spatial characteristics are learned using convolution and recurrent layers. An accuracy of $98\%$ was reached in detecting accidents in publicly available road accident records, suggesting a great capacity for detection regardless of the road surface.

Object Detection

The detection of objects requires an intelligent platform to detect them objects effectively in the transport system (Zhang et al., 2021). The trajectory design using multilayer spatial networks achieves optimal solutions (Jia et al., 2022). Image detection requires intelligent semantic technology to detect objects in the urban background perfectly (Zhou et al., 2021). Recently digital twin algorithms are widely used in VANET for secure energy optimization (Chen, 2022). Due to improper traffic, the management environment is highly polluted and brings harm to human life (Xiao et al., 2020). Routing and navigation are key requirements of smart vehicle technologies (Xiao et al., 2021; Sun et al., 2021). For guiding the path, the Saccades Recommendation is used by drivers in urban areas (Xu et al., 2022). Recent developments in technology help blind people or sight-impaired people in traffic analysis and the selection of the optimal path (Xu et al., 2021; Ren et al., 2022). The train is the biggest network and the transport system needs the utmost intelligence to control the traffic or delay in travel (Yin et al., 2022; Xiao et al., 2022; Zhuang et al., 2022). The features extracted for the traffic system are processed to wavelet filters to remove unwanted noises in features (Liu et al., 2022b). Wireless data transmission or wide band coverage become a major issue in traffic analysis applications (Liu et al., 2022a; Feng et al., 2022). Location tracking for logistics applications uses supervised learning techniques for object detection (Dai et al., 2022; Xie & Sun, 2022). The wireless communication system plays a vital role in transferring data in a secure and quick manner (Yang et al., 2022). Path planning using graph networks (Xie et al., 2022) helps in tracing an optimal path from source to destination (Cao et al., 2022). A deep learning-based classification algorithm for classifying roadside data is discussed in the article (Shen et al., 2020; Zheng et al., 2022b). This research collects roadside data and performs analysis for identifying traffic (Zheng et al., 2022a; Ban et al., 2022). The detection of objects (Li et al., 2021) using transfer learning can achieve an accuracy rate of more than 95. The fusion of sensors helps to predict the objects or collision objects in a better way (Du et al., 2021). The issue only affects car accidents; motorcycles, bikes, and walkers are not included.

Existing Work Based on IoT and Various Transport Methods

Proposed Methodology

The proposed method uses an Adaptive Traffic Management system (ATM) with an accident alert sound system (AALS) and Secure Early Traffic-Related EveNt Detection (SEE-TREND) for a secure and intelligent transport system. The proposed method consists of four layers for developing an innovative and safe transport system. Initially, the application layer monitors the vehicle’s location and image tracking, and then the accident alert sound system tracks the accident. The next layer is the service layer which gathers data, and then the collected data is pre-processed. The third layer is the network layer which is used for data communication. In this layer, Secure Early Traffic-Related EveNt Detection is used for transferring vehicle data securely. Finally, the sensing layer which collects the data with the help of sensors.

Application Layer

In this layer, the vehicle location, vehicle image tracking, and then the attack detection phase is implemented.

Tracking the Location of Vehicles

The suggested ATM-ALTREND system aids in selecting routes with greater accuracy. The lower limit precision level of the test is used to evaluate the quality of the model. If the suggested model produces the lower bound with the appropriate level of precision, good, efficient pathways exist, and all other, less effective communication channels have been eliminated. The set of essential ways for efficient vehicle positioning is expanded if there need to be more routes and the lower bound is higher than the anticipated precision rate. Information is gathered in the initial stage utilizing the sensors and photography equipment. Preparing data after sensors or cameras have captured it is an essential part of ITM. When preprocessing data, missing value estimate techniques are employed. The gathered data are processed using the processing method, and the dataset is then trained using the training method. Traffic information and the precise position of the vehicle are gathered.

Feature Clustering

After tracking the locations of the vehicle, the feature clustering process is held. To avoid feature clustering, a graph is created. The nodes in the network (feature groups) reflect vehicle sightings, the edges represent connections between path clusters, and the nodes represent feature routes. For each trait and characteristic at a given time ($TI_i$):

• If the estimated total individual movement is sufficiently large, features discovered at a time interval ($TI<em>i$) for the frame ($FR</em>i$) are selected and tracked for a threshold number of frames. A Euclidean distance minimum connects almost all newly produced features that are extracted to the currently recorded characteristics.

• The higher and lower limit intervals are updated along with an approximation of the distance ($Dist<em>{i;j}$) between all currently monitored sets of connected functionality ($LIf</em>{i;j}$). The amount of the features separation threshold is represented by the $D_{fseg}$. The attributes of the connected automobiles are described in (1).

$Max<em>{TI</em>i}d<em>{ij}(\ ) - TI</em>i \leq Min<em>{TI</em>i}d<em>{ij}(\ ) - TI</em>i \leq D_{fseg}$

• The related aspects of the graph are identified. Every combination of feature pathways that make up a related component, or vehicle observation, represents one. Let us assume that a component no longer has any functionality that is documented. The properties of the vehicle assumption (velocity vectors, centroid location, and vehicle type) are estimated after the characteristics are removed from the graphs.

Accident Detection Using Accident Alert Sound System

By using intelligent traffic control during Multiple Vehicle Collisions (MVCs), an ITM system can lower the likelihood of accidents and the number of unintentional fatalities (MVCs). To prevent MVCs, a system operating outside the car is required for accident detection and alarm generation for approaching vehicles. Various types of sensors and actuators are fitted in smart roads (SRs) for the automated detection of accidents. Nodes in SRs are spaced roughly 50 m apart. One can use their previously saved placements since these nodes are attached to the road’s side of the road. To facilitate the quickest rescue effort and minimize damage, a node that notices an accident notifies its previously saved position to an EOC (Emergency Operation Centre). An AALS alert system that warns drivers of oncoming cars or accidents is the critical component of the SR. The AALS is explicitly created for EVs and EVs to prevent MVCs. It is challenging to outfit all vehicles with the same communication and preventive system. The method employs a siren and a golden yellow blinking light to warn passing drivers of an accident. A loud siren is another alternative method of warning motorists, allowing them to hear it and take action to prevent an MVC, particularly in BWC. These notifications are produced automatically. Both sides of the street will have the proposed system implemented.

The primary purposes of the AALS technology are accident detection coming from external cars (i.e., from the side of the road) and accident notification to oncoming traffic via light and sound flashing. Certain things happen whenever an accident occurs. For instance, screeching noises are made when the brakes are applied quickly, and loud noises can be detected far away when a car collides with a different one. Glass shattering makes an audible noise; a burning car raises the atmosphere’s temperature and releases smog; if a vehicle immediately stops in the center of the road, it creates a hazard for other vehicles. A flawless accident detection system is available by considering all of these elements.

The microcontroller receives input from the microphone, infrared, smoke, and HC-12 sensors. To find fires, a smoke sensor is employed. The microphones capture the audio and transmit it to the microcontroller that analyzes the levels with a predetermined threshold when it comes from a vehicle collision or hitting an item. If the level exceeds the threshold, an accident is declared. Therefore, the microcontroller dismisses the sound if it is not louder than the threshold. The IR sensors sense objects on the roadway concurrently in the interim. The microcontroller analyses the IR signal to see if it falls below the predetermined threshold and if more than 5 min have passed.

Vehicle Image Processing in IMT

This technology uses pictures and electronic systems built into the roadways first to identify the automobiles. In addition to the traffic signal and sensors, the webcam will also be put in place. This will record visual data trends. Anomaly detection is the best choice to control the change in the status of the traffic light. A green signal over an empty roadway can decrease lost time and lessen traffic jams. Additionally, as it uses actual traffic-image data, it is more reliable at predicting the presence of vehicles. More significant than all other designs that rely on recognizing the vehicular surface material, it examines the usefulness and procedures.

Vehicle Communication

The sensors identify the car and its communication devices and continuously track its position in the traffic flow. Information is sent and shared amongst vehicles using IoT sensor devices in a manner that aids in reducing traffic and assuring travel safety. The platform is designed to alert users in advance of vehicular and dangerous driving situations, effectively handle accidents and fatalities, and address problems with data to be recognized by motorists on the road for safe traveling. Additionally, it is essential to communicate the information to the drivers by obtaining an astronomically large number of previously accurate statistics depending on the present traffic circumstances. In VANET, communication is crucial and is facilitated through the roadside unit (RSU).

Secure Traffic Data Using Secure Early Traffic-Related EveNt Detection

SEE-TREND gathers traffic information from passing vehicles. Since the cars in localized cohorts experience the same traffic conditions, this information correlates. As a result, particularly in areas of heavy traffic, it is optional to gather data from every passing car; instead, we will focus on investigating the issue of probability collecting data. Suppose that its Traffic Monitoring Unit (TMU) throws a biased coin that comes up tails with probability $p$ and captures information from a particular moving car with probability $p$ rather than gathering information from every passing vehicle. Assuming also that, for such application-dependent value >0, the data gathered from $k$ cars results in the accurate detection of a traffic-related incident with probability $1 - e^k$. Let $AEn$ become the instance when based on the aforementioned hypotheses, $n$ passing cars are sufficient to accurately detect a traffic occurrence. Let $Hr serve$ as the random vector that determines the number of the $n$ passing cars that have contributed traffic data. By training for the upcoming $Hr$, we now

$Pr[\leq AEn \leq] = \sum_{k=0}^{n} Pr[AEnjHr = k Pr[Hr = k]$

$= \sum_{k=0}^{n} {\binom{n}{k}} p^k(1-p)^{n-k} (1 - e^{-ak})$

$= \sum<em>{k=0}^{n} {\binom{n}{k}} p^k(1-p)^{n-k} - \sum</em>{k=0}^{n} {\binom{n}{k}} (pe^{-a})^k(1-p)^{n-k}$

$= 1 - [1-p(1-e^{-a})]^n$

In addition, the vehicle arrival rate, which in turn is based on traffic flow, affects the value of $n$. A specific TMU can calculate the number $n$ of entries per unit of time if it is informed of the frequency of traffic flow. This will then make it possible to calculate $p$. If the traffic flow intensity allows, the TMU can consciously choose not to gather information from a certain number of vehicles. This will decrease TMU power usage without affecting the time it takes to notice a problem connected to traffic.

SEE-TREND Security Solutions Implementation

The SEE-TREND was created to permit the anonym collection of data while guarding against impersonation attempts. Security from privacy attacks is included in SEE-TREND. Even from a position along the roadside, the attacker cannot understand the handshaking technique between a TMU and a moving car since it uses encrypted data.

Establishing frequencies for communication between the TMU and the moving vehicles is an essential component of the handshake. By doing this, the opponent will be unable to understand the future data exchange. Among the main challenges of SEE-TREND is minimizing the effects of DoS assaults, which can take many different forms. In one type of DoS attack, two opponents, X and Y, acting simultaneously, each placed by the edge of the highway near neighboring TMUs A and B, try the following technique: The keys for the new vehicle are all picked up by X and sent to Y via TMU A. Y will then communicate to B while posing as a legitimate vehicle. Whenever the vehicles that provided the key try to connect TMU B, the request would be refused because the opponent has already utilized the one-time key, which TMU B will recognize as fake (as the time stamps do not match). The fundamental question, in this case, is how SEE-functioning TRENDs may be maintained when a TMU is shut off due to a failure or energy outage.

As seen in Fig. 4, the goal is to create a straightforward clustering algorithm on the collection of TMUs by clustering together $k$, (k > 2), neighbouring TMUs. The clustering factor, or $k$, will be used from now on. Every TMU in this clustering method only joins one cluster, leading to fragmented clusters as a consequence. According to the theory, communications can transmit successfully from end to end in accordance with the SEE-TREND semantics as soon as each cluster has an active TMU. The TMUs can be clustered either statically or dynamically. The likelihood that, under incident X, SEE-TREND would failure to transmit messages end-to-end is a solid indicator of fault tolerance.

\sum{k=i}^{n} {\binom{n}{ki}} p^{ki} (1-p^{k})^{n-ki} = 1 - (1-p^k)^ {n\k} = 1 - e^{\frac{-np^k}{k}}

Mathematical Model for ATM-ALTREND

The "platoon-based traffic flow" is the foundation of the suggested ATM-ALTREND model. A platoon is typically thought of as a group of cars traveling side by side, either voluntarily or involuntarily. Each vehicle is predicted to receive an impartially required power $V$ in the suggested ATM-ALTREND mathematical formula, and the traffic variance, $FT (VE)$, is given. Automobiles travel continuously if they close the distance to the traffic in front of them. Once it has caught up, the object immediately slows down to match the speed of the moving object and follows it while maintaining steady progress. The road traffic in terms of available capacity vs. time is depicted in Fig. 5. The road lane’s length is denoted by the letter $LE$. $TVIn$ denotes the time of entry, $TVout$ denotes the time of exit for the vehicle, $VEi$, and ($TVhi$) denotes the time of headway for the vehicle, $VEi$ indicated in Eq. (4).

$TVout = TVout_i (for i = 0)$

$Max(TVout<em>i; TVout</em>{i-1} + TVhi (for i = 1; 2; … ; n) )$

In Eq. (5) it can be used to compute the time-out $TVout$. There are no cars near the boundary configuration when $i = 0$.

$TVout = fTVin_i + (L=VEi)g$

Let $PV (EVs)$ represent the probabilistic likelihood of an $N$-car platoon attempting to produce an $EV$ event. Therefore, $EV$ is the occurrence, and $PV (N|EVs)$ is the probabilistic likelihood that a platoon of $N$ cars will cause event $EVs$ to occur. At the conclusion of the evaluation phases, the likelihood $PVPrimaryVE (N)$ that a particular new car will be the lead vehicle in a platoon of $n$ cars given in Eq. (6).

$PVPrimaryVE() = N - lim<em>{0\leq1} PV</em>EVs() - PV N() - jEVs FVE(PV)dv$

Machine Learning Models for ATM-ALTREND

Next the machine learning model is used for training and testing the traffic data. The significant infrastructure required for the development of the smart city, intelligent transport system is a hotly debated field of study. Additionally, it employs IoT and ML technology to resolve disputes in a positive manner. The dataset contains all relevant information about the traffic environment. It contains information on the car, the road, and the traffic. The following steps are involved in this process.

• Step 1: The dataset contains vehicle data, road and traffic length and accident locations are collected initially.

• Step 2: The collected dataset is pre-processed and then the pre-processed data is given into training.

• Step 3: The trained dataset is given into Machine learning method. Here machine learning based DBSCAN clustering method is used.

• Step 4: Next the trained model and testing dataset is evaluated. Finally the desired outcome is produced.

Result Analysis

A MATLAB simulator was used to construct the suggested ATM-ALTREND system. The suggested system makes use of the three main components (vehicles, infrastructures, and occurrences) listed in Table 2.

Different traffic situations for linked autonomous vehicles (LAVs) were developed in order to verify the effectiveness and reliability of the suggested ATM system.

• (a) Only when connected to LAVs;

• (b) where few non-LAVs are present;

• (c) where both LAVs and non-LAVs are traveling.

Three Scenarios

The proposed method evaluates three scenarios such as only with LAVs, where only with non-LAVs, and where LAVs and non-linked both types of vehicles are moving.

Only with LAVs

It is the first situation that solely takes LAVs into account. In this case, the traffic is primarily divided into two groups by the intelligent traffic-management systems. The control segment (CS) is the first, and merging segmentation is the second (MS). Control unit (CU), a component of the CS, aids in communication with LAVs.

Where Only with Non-LAVs

Evaluations are required to confirm the viability of suggested ATM-ALTREND techniques. In order to enable users to evaluate multiple viewpoints, a traffic virtual environment system must be simple to adjust to different traffic scenarios. On the basis of vehicular modeling, a baseline sequence of occurrences is created and assessed, with just fixed-cycle traffic lights monitoring non-LAVs.

In Which Both Kinds of Vehicles—LAVs and Non-Linked—Are Traveling

The mixed-traffic scenario, in which LAVs and non-LAVs use the same roads, should be seen as a major obstacle to the widespread adoption of automated cars. The suggested course of action in this circumstance is tested using system model control approaches. In order to execute the planned ATM-ALTREND system, IoT sensors will gather real- time data on vehicular traffic and driving conditions and store it in a data center. Big data processing methods will be used to analyze the real-time data in the following phase. The final stage will rely on machine learning techniques to analyze the data. We presume that the wireless transmission component that communicates with the RSU in the suggested ATM-ALTREND system is fully installed in motor vehicles. Situated mostly on the highway, (Fig. 5) is designed to transmit substantial traffic information with other moving vehicles.

Simulation Environment

The simulated results show that the suggested ATM-ALTREND system works better in terms of the packet delivery ratio, bandwidth, and time delay than the traditional traffic- management system (Lilhore et al., 2022). This functions best when latency and traffic conditions (as in scenarios 1 to 3) change. A 5,000 m road, 100 vehicles, and four different traveling directions (Channels 1: East to West, 2: North to West, 3: North to East, and 4: East to North) were used in the test, which produced the findings. Traffic moves in both outflow and inflow directions in every simulation. The random model is the type of model. The likelihood of inbound and outbound traffic is $0.52 \%$ and $0.43 \%$, respectively.

Scenario 1

In this scenario, it is only with Linked Automated Vehicles (LAV). The experiment evaluates the traffic congestion ratio, time means speed, harmonic mean and jam ratio during whole time.

The proposed method has better performance in experimental results. It reduces the traffic congestion, jam ratio during whole time. Also it reduces the time mean speed and harmonic mean.

Scenario 2

In this scenario it belongs to non-LAVs, the experiments were carried out by using the following metrics such as traffic congestion ratio, time mean speed, harmonic mean, and jam ratio during the whole time.

The proposed method achieves better performance in all the parameters used for the transport system.

Scenario 3

In this scenario it belongs to both kinds of vehicles—LAVs and non-linked—are traveling, the experiments were carried out by using the following metrics such as traffic congestion ratio, time mean speed, harmonic mean, and jam ratio during whole time.

The proposed method achieves better performance in all the parameters used for transport system.

Clustering DBSCAN and Machine Learning Method

A traffic simulator without collisions is the MATLAB simulator. A DBSCAN technique based on machine learning will be used to find the accident in the following stage. The clustering results of DBSCAN and ML techniques for accident avoidance are shown in Table. Each vehicle in the simulation must come to a complete stop at a road view every 95 s. The clustering method uses two types of clusters normal and anomaly. The proposed method ATM- ALTREND detects the normal and anomaly cluster type among the vehicle count.

Conclusion

In this work, an Adaptive Traffic Management system (ATM) with an accident alert sound system (AALS) is used to manage traffic congestion and detects an accident. The intelligent transport system alerts vehicles during accidents and minimizes collisions during traffic. Secure Early Traffic-Related EveNt Detection (SEE-TREND) is used for secure traffic data transmission. The ATM-ALTREND model that has been developed provides:

• On-site auto operations.

• Smart parking.

• The application of traffic-management strategies for the creation of an intelligent transportation system.

The plan aids in tracking vehicle travel, allowing for the analysis of traffic in a particular area. The SEE-TREND minimizes intrusion and secures the traffic data during data transmission. In this work, three scenarios are when connected to LAVs, where few non- LAVs are present, and where both LAVs and Non-LAVs are traveling. In these three scenarios, the proposed ATM-ALTREND outperforms better than the existing work.

With the use of vehicle locations and average speeds, we were able to demonstrate that traffic activity could be assessed convincingly. Drivers closest to the accident scene may also view unusual roadway incidents as a potential threat. It was discovered that the suggested ATM-ALTREND system outperformed the currently used traditional processes.