Data shows 3500 automobile accidents each day worldwide. According to the UN, India has the most road fatalities. The drivers' lack of information contributed to this tragedy^{ }
To establish a new hybrid deep learning approach for predicting the situation of heavy traffic.
To select the Optimized BiLSTM (OBidirectional Long short Memory)to forecast traffic congestion.
To enhance the traffic detection accuracy, the weight of BiLSTM is tuned using the new hybrid optimization model.
The remainder section is ordered as: Section 2 of the paper examine about the existing works that has been accomplished. Section 3 portrays information on the proposed methodology for Traffic Congestion Control System For VANET. The results gained with the proposed model is discussed thoroughly in Section 4. In Section 5, this paper is concluded.
In 2018, Jain et al.^{15} have suggested city lane and crosssectional simulationto accomplish the goal of dominant vehicle mobility. It displays actual communication between cars and between cars and traffic infrastructure. Network Simulator 2 was used to conduct the experiment. 3 units all needed to be modelled for the implementation. In the simulation, factors like traffic volumeand packet loss were led intoaccount. These variables guarantee effective signaltosignal communication. As the vehicles would be not reported to access the intersection area and given data pertaining to other vehicles, this improves traffic control and road safety.
In 2019, Liu et al.^{ }
In 2019, Mohanty et al.^{ }
In 2018, Sharma et al.^{18} have created and validated ESPM framework to forecast the likelihood that an accident wasoccurred on an Indian fourlane highway. To foresee an emergency situation in advance was the main goal of ESPM. So, it works to reduce the number of fatalities and injuryinduced by traffic collision. Three stages—reporting, monitoring, and prediction—was involved in ESPM's emergency situation prediction process.
In 2017, Ravikumar et al.^{19} have enhanced congestion control by using heuristic techniques to decrease traffic announcement channels while taking into account the reliability of submission supplies in VANETs. The simulation results have showed that metaheuristic techniques significantly outperform other blocking control procedures in VANETs.
In 2020, Abdelatif et al.^{ }
In 2017, Mallah et al.^{ }
In 2020, Choe et al.^{22} have suggested a cooperative RLpowered intelligent channel access algorithm in which vehicles fully decentralise channel admittance coordination. In order to improve the V2V safety broadcast in congested, infrastructurefree VANETs, also took into consideration an appropriate interaction scheme between vehicles. Agreeing to different levels of traffic congestion, presented evaluation results from indepth simulations. Additionally, the algorithm meets both the short and longrun communication fairness requirements as well as the wait time essential of VANET secure application.
In 2018, Ullah et al.^{23} have presented urgency message dispersalschedules that was based on VANET situation of preventing congestion and the use of vehicular Fog computing. A VANET architecture that usedFoG assistance was investigated in order to competentlydeal with message congestion scenarios. Also outlined a taxonomy of strategies for preventing message bottleneck. To highlight the favors and shortcomings of different congestion avoidance strategies, included a comparison discussion.
In 2021, Kothai et al.



Ngo et al. ^{2} 
enhancing the safety of intersections on roads 
Minimum waiting time per vehicle performance metric is not improved using a genetic algorithmic approach. 
Zheng et al. ^{5} 
to install the bare minimum of RSUs necessary to distinguish and cover all rushhour traffic. 
Vehicle flow's spatiotemporal characteristics are not examined. 
D'Andrea et al. ^{7} 
realtime GPS monitoring for traffic jams and incidents 
A dynamic routing service is not integrated with the system. 
Elhoseny et al. ^{13} 
Utilising a Clustering Model for Energy Efficient Optimal Routing in VANET Communication 
low cost, energy efficiency remains unchanged. 
Qureshi et al. ^{14} 
A dynamic congestion monitoring programme for vehicular ad hoc network safety applications 
Techniques for probabilistic traffic prediction are not investigated. 
Jobaeret al. ^{8} 
VANETBased UAVhelped Hybrid Scheme for safety on city roads 
Amount of RSU is not reduced. 
Choe et al. ^{22} 
Congested vehicular networks require a robust channel access method that uses cooperative reinforcement learning. 
Performance is not improved. 
Kothaiet al. ^{24} 
Prediction of Severe Traffic Congestion in Smart Cities 
• No larger datasets are available. • Prediction accuracy is not improved. 
Step1: The data is collected from: https://www.kaggle.com/datasets/arashnic/roadtraficdataset?select=region_traffic.csv
Step 2: Then, the data features like statistical features, higher order statistical features, correlationbased features and database features is extracted from the collected data.
Step 3: The traffic congestion is predicted using the new hybrid deep learning approach (proposed) that includes the Recurrent capsule networks (CapsRNN)
In this research work, information is collected from the Traffic Road Dataset. For each junctiontojunction link on the motorway and 'A' road network, as well as for few tiny roads in Great Britain, road traffic open data provides streetlevel information. Vehicle miles, which combine the quantity of vehicles present on the road and the distance they travel, are the most common units of measurement for annual statistics. Details from about 8,000 roadside 12hour manual counts, continuously updated traffic counters knowledge, and information on road lengths are used to create annual traffic statistics. Every 10 years or so, the road traffic information group conducts asmall route traffic benchmarking exercise with the goal of increasing the precision of traffic projections for small paths. The exercise's findings from 2018 to 2019 have been made public, and the minor road has been modified.
•
Where,
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
In this research work, the extracted features are predicted using the new hybrid deep learning approach (proposed) that includes the Recurrent capsule networks (CapsRNN), Fuzzy Interface System (FIS) and Optimized BiLSTM. Novel hybrid deep learning approach is trained using the extracted statistical features. In addition, to enhance the traffic detection accuracy, the weight of BiLSTM is tuned using RSOA. It is the combination of the standard SWOA and RO. The outcome from Recurrent capsule networks (CapsRNN) and Optimized BiLSTM is fused together, and it is fed as input to Fuzzy Interface System (FIS), which makes the final detection regarding the network congestion. The data acquired from FIS is stored in the cloud.
The input layer, recurrent layer, and capsule layer are the three main parts that the network integrates.
In a sentence related to coronary arteriography
The LSTM is an RNN variant that uses a gated memory cell to occupy extendedrange dependencies in the data and has the capability to prevent the rapidly increasing issues that are present in conventional RNNs.
where
Because of the present condition
Five LSTMs were used to handle the five segments of the sentence because the two entities in the sentence can be divided into that many parts. Particularly, the context is less impact on relation classification and that there is a greater distance. However, BiLSTMs are used for the three middle categories lefttoright
The first application of the capsule layer is digit recognition. A capsule is a group of neurons whose activity vector illustrates the instantiation parameters of a particular kind of connection which uses only a neuron to represent the classification probability via a sigmoid function or a softmax function: The probability that the corresponding relation exists is represented by the length of the activity vector, and different vector orientations can designate various cases under the relation, giving the capsule a greater capacity for expression.In order to ensure that short vectors are lowered to almost zero length and long vectors are decreased to a length just underneath 1, which takes into account the fact that the length of a capsule is exploited as the chance of a relationship occurring.
Where
As shown in
A fuzzy logic system is composed of a fuzzy inference system, which utilizes fuzzy set theory, IFTHEN rules, and fuzzy reasoning to determine the output that corresponds to crisp inputs.
• Obtain precise value from the action
• Uses the fuzzy membership function to transform the crisp value into a fuzzy value.
• Utilise the IFTHEN rules in the fuzzy rule base to produce fuzzy results.
• Apply some Methods used to convert fuzzy output into a precise/crisp value.
• A detailed understanding of Fuzzy's construction is possible thanks to its five functional building blocks.
• The fuzzy IFTHEN rules make up the set of rules.
• The functions defining the membership of fuzzy sets employed in the fuzzy rules are defined in the database.
• The decisionmaking unit applies the fuzzy rules in its operations.
• Any crisp quantity can be transformed into a fuzzy quantity with the aid of the fuzzification interface unit.
• Any fuzzy quantity can be transformed into a crisp quantity using the defuzzification interface unit.
• These steps can be used to breakdown the fuzzy inference system's overall operation.
• The various applications of the fuzzification techniques is supported by the fuzzification unit.
• When the crisp input is changed to a fuzzy input, a rulebased database's knowledge base is formed by assembling a collection of rules.
• The fuzzy input from the defuzzification unit is finally be transformed into a crisp output.
A weight function is used in a BiLSTM to determine the relative weights of each input feature and hidden state at each time step.A group of learnable parameters that are optimised during training define the weight function. These parameters specify the contribution of each input feature and hidden state to the network's output.
Evaluate the fitness of each search agent using Eq. (15).
Where,
During the exploration space, the RO is used to explore the search space by simulating the movement of rays of light. The search agent moves randomly within their neighbourhoods in a Zigzag pattern to cover the search space thoroughly. This movement can be given as:
During the exploitation phase, the SmallWorld Optimization Algorithm is used to exploit the best search agents found during the exploration phase. The search agents advance the guidance of the best search agent within their neighbourhoods to converge towards the optimal solution.The movement of each search agent is expressed using the following mathematical equation:
Update the fitness and neighborhood structure: Evaluate the fitness of each search agent using the objective function and update the neighbourhood structure using the SmallWorld Optimization Algorithm.
Termination: Terminate the algorithm if a condition for stopping is fulfilled (e.g., supreme number of iterations is accomplished, the fitness of the best search agent does not improve over a certain number of iterations, etc.). Otherwise, return to step 4.
The suggested model has been executed using MATLAB. The intended model has been analysed in terms of Accuracy, Precision, FMeasure and Recall.
Accuracy, Fmeasure, precision, Recall is used as comparison metrics for performance.
Accuracy is the degree of closeness between a measurement and its true value. A classification model's performance can be assessed using the metric of accuracy. According to Eq. (18), accuracy is the prediction model's percentage for both the number of values that were correctly predicted and the overall number of predicted values.
Precision refers to the amount of information that is conveyed by a number in terms of its digits. The model's precision measures how accurate it is, or how many of the positive predictions have turned out true. The mathematical expression is shown in Eq. (19)
The Fmeasure can be viewed as a compromise between recall and precision. F1 Score seeks to strike a balance in recall and precision. The mathematical expression of FMeasure is shown in Eq. (20)
The ability of a model to find all the relevant cases within a data set. Recall measures how many Actual Positives the model identified properly. The mathematical expression is shown in Eq. (21)
A new hybrid deep learningbased model is presented to accurately predict traffic congestion on roads. To reduce traffic congestion in connected cities, this research proposed a new traffic congestion control system using hybrid deep learning approach techniques that gather traffic data on available routes. Training and testing are the two main phases of the suggested model. The data are collected from kaggele about 8,000 roadside 12hour manual counts. The Road Traffic Dataset is where the information is found. The gathered data are extracted using database features, higher order statistical features, correlationbased features, statistical features. From the extracted data, a new hybrid deep learning approach using recurrent capsule networks, fuzzy interface systems, and optimised biLSTM is used to predict traffic congestion. The newly developed hybrid deep learning approach (RSOA) is trained using the statistical features that were extracted. Additionally, the weight of the BiLSTM is adjusted using the new hybrid optimisation model to increase the accuracy of traffic detection. The proposed model TCCHDL has been analysed in terms of Accuracy, Precision, FMeasure and Recall with the standard.