A Novel Method to Improve the Prediction of Vehicle Numbers

Involved in Crashes at Rural Areas Using Reinforcement Learning

Models

Marthandan T.

, Veena P.

, Jeyabarath R.

, Pragadeesh G.

, Prithiviraj S.

and Ragavendhra A.

Department of ECE, K.S.R College of Engineering, Tiruchengode, Namakkal, Tamil Nadu, India

Department of EEE, Professor, K.S.R College of Engineering, Tiruchengode, Namakkal, Tamil Nadu, India

Department of ECE, K.S.R. Institute for Engineering and Technology, Tiruchengode, Namakkal, Tamil Nadu, India

Keywords: Reinforcement Learning (RL), Prediction Accuracy, Vehicle Crashes, Real‑Time Data, Traffic Conditions,

Rural Areas, Statistical Model, Crash Prediction, Environmental Factors, Road Safety.

Abstract: Aim: The current study aims to design a new approach to enhance the prediction of the number of vehicles

involved in crashes in rural areas using reinforcement learning models. Materials and Methods: Two groups

were compared, where Group 1 is a traditional machine learning approaches how much vehicles involved in

accidents based on historical crash data; Group 2 is a reinforcement learning (RL) model with a crash data

driven model, which integrates crash data and can responds to real-time traffic as well as environmental

factors for feedback, and can dynamically adjust prediction. Result: The system shows improved prediction

accuracy compared to traditional Conventional model. The mean accuracy of the RL model is 95.2% while

the mean of the comparative model is lower than the RL model and it is 88.9%. This increase in accuracy was

statistically significant (p =.042), as verified by the independent samples test. Conclusion: This study

demonstrates that the use of an RL-based prediction model yields reliable and higher performance in

predicting the numbers of vehicles involved in crash events at rural locations. Also, this combination offers

more plentiful and evolving detention action plans providing greater road safety.

1 INTRODUCTION

According to the World Health Organization traffic

accidents are one of the leading causes of death,

especially in rural areas where resources and response

times may be lower. Such prediction system can go a

long way in enhancing the safety as well as preventive

measures for vehicle crashes in these regions. Jaradat

S, et al., 2025 This study explores a new approach for

predicting the number of vehicles involved in crashes

(IVC) at rural areas based on reinforcement learning

models.

Historically, machine learning has been used based

on historical crash data that studied dynamic factors

such as traffic conditions, environmental influencing

factors in real time. Zhang G., et al, 2024, The

proposed system utilizes the combination of

reinforcement learning with real-time data to enhance

the prediction accuracy, allowing timely warning

alerts for preventive action. In this paper, we will

introduce an RL-based prediction model that

considers various data points analyzing crash vehicle

count. Anand Kumar G., et al., 2025, By integrating

real-time data about the surrounding environment and

traffic, the self-parking system can adapt to changing

conditions, improving its performance. Vinoth B., et

al, 2025 This novel approach tries to deliver more

accurate forecasts, which allows public agencies to

allocate resources efficiently and enhance road safety

in rural parts.

Moreover, Zhang C., et al, 2025., demonstrated the

feasibility of applying reinforcement learning in

various predictive scenarios, highlighting the potential

of this approach for real-world applications. In this

context, however, the technologies bear mainly the

idea to improve the life quality and safety of those who

might remain in rural areas, while reducing the harm

and impact of vehicle crashes. A new predictive

system developed here via reinforcement learning

models shows it is possible to close the gap between

existing predictive tools and the expectations of real

traffic control design.

690

T., M., P., V., R., J., G., P., S., P. and A., R.

A Novel Method to Improve the Prediction of Vehicle Numbers Involved in Crashes at Rural Areas Using Reinforcement Learning Models.

DOI: 10.5220/0013919100004919

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 1st International Conference on Research and Development in Information, Communication, and Computing Technologies (ICRDICCT‘25 2025) - Volume 4, pages

690-696

ISBN: 978-989-758-777-1

2 RELATED WORKS

Research spanning the previous decade has

thoroughly investigated over 2,800 studies of

technological improvements to estimating vehicle

counts involved in crashes, within predominantly

rural settings. Many of these studies have focused on

traditional statistical models as well as machine

learning techniques, recently reinforcement learning

(RL) has developed into a potential tool in this area.

KML-KYW: Traditional models have usually

depended on historical crash data, while RL models

use the real-time traffic and environmental data to

better prediction performance. All the references are

taken from very well-known IEEE terms, journals,

and existing research papers.

Karanikas N., et al., 2020., A major study used a

standard statistical model based upon historical crash

data to estimate numbers of vehicles involved in

crashes. Poonia RC., et al., 2022., While widely

accepted, this method typically neglected to consider

dynamic aspects such as current traffic states and

other environmental effects. Dhinesh Kumar R., et al,

2025 On the contrary, real-time data has been used in

recent work to show that reinforcement learning

models are more effective crash predictors. Khan SS.,

et al., 2025 The RL approach outperformed

traditional models with a mean prediction accuracy of

87.8% to 95.2%, compared to the traditional model's

82.5% to 88.9%. Pusuluri VL., et al., 2024 A second

highlight paper describes deploying an RL-based

system that dynamically tuned its predictions based

on contemporary traffic and environmental

conditions. It used multiple data sources like car

speed, climatic conditions and mode of road for better

prediction. Reddy JS., et al., 2025 The results showed

a significant performance gain in prediction, with a

7% improvement in accuracy compared to traditional

models. Independent samples test indicated that this

increase in accuracy was truly significant at 0.042.

Research from November 2020: We only explored

moving common RL models to different predictions

below, but our set of predictions also mapped to

different voltages which were more relevant to

different scenarios Qawasmeh BS., 2024. Shamim

Kaiser M., et al, 2021 These studies highlighted the

need for real-time data to improve the accuracy of

predictions and warn in advance to enable preventive

measures. The results indicate that RL based models

have the potential to outperform conventional ones,

marking an important advancement in the prediction

and prevention of crashing.

In summary, the present study demonstrates that

reinforcement learning models can enhance the

prediction of vehicle volumes of crashes in rural

locations. By incorporating real-time data, RL models

offer more precise and current predictions, thereby

improving road safety interventions. We propose an

improved prediction system where reinforcement

learning models are adopted to improve prediction

and thus contributing in limiting the gap between

modern predictive tools and the requirements of

practical traffic operations.

3 MATERIALS AND METHODS

Using the real time traffic and environment data, a

new method was developed and implemented

stepwise to model the number of vehicles involved in

crashes in rural settings in KSRIET IOT laboratory.

The method utilizes reinforcement learning (RL)

models to optimize prediction accuracy and enable

timely alerts for preventive actions. Combining RL

with real-time data makes the system adaptive,

leading to efficient prediction and management of

vehicle crashes in dynamic rural environment.

Figure 1: Flow Cycle of the Reinforcement Learning-

Based Crash Prediction Mode.

Group 1: The conventional techniques of vehicle

crash prediction used historical crash data & classical

statistical models. After that, these models were re-

evaluated on 100 crash incidents with a mean

processing time of 21.1ms and an accuracy of around

85%. Figure 1. The complete system is shown in

A Novel Method to Improve the Prediction of Vehicle Numbers Involved in Crashes at Rural Areas Using Reinforcement Learning Models

691

Figure 7. Traditional methods were focused on

parameters like historical data analysis and pattern

recognition but lacked real-time adaptiveness and had

low accuracy under dynamic operational condition.

Group 2: Reinforcement learning models are

employed by this system, enabling its real-time

predictive adjustments according to both current

traffic and environmental data. In order to test the RL

model, it was trained with a dataset of 200 crash

incidents, which averaged 12.1ms to 17.3ms of

processing time, and provided an improvement in

prediction accuracy with average results of

approximately 87.8% to 95.2%. Relying on dynamic

prediction and real time responsiveness, RL with

respectable performance, overcoming the limitations

of less precise and adaptive to rural traffic changes.

Structured Flow of the Novel Prediction System for

Vehicle Crashes in Rural Areas the RL model and

data collection modules are the ones initializing the

systems in this manner.

This is given by the absence of the traffic and

environmental data in a time-step up to the RL model,

as the recent time v or within T time v have been

continuously captured in real time and again

configured as an input for the RL model to predict a

correct data. This data must be fed into an RL model

that will learn to predict the number of vehicles

involved in potential crashes. In case of prediction for

critical condition, alerts are raised instantly to notify

the concerned authorities for preventive measures. As

new data is ingested, updates the predictions made

about where, how and when crashes would take place

in rural region of the nation.

4 STATISTICAL ANALYSIS

SPSS version 26.0 was used for the statistical analysis

of data collected from parameters such as prediction

accuracy (%), F1-score, and processing time

(milliseconds). The independent sample t-test and

group statistics were calculated using SPSS software.

The reinforcement learning (RL) models and

traditional statistical models were considered

independent variables, while prediction accuracy

(%), F1-score, and processing time (milliseconds)

were dependent variables.

5 RESULT

The Performance of the novel method to improve the

Prediction of Vehicle numbers involved in Crashes at

rural areas using Reinforcement Learning Models.

Table 1: Comparative Performance Metrics of RL Model and Conventional Model on MRI Image Dataset.

Sl.

Image

Precision

(%)

RL Model

Precision (%)

Conventional

Model

Recall (%)

RL Model

Recall (%)

Convention

al Model

F1-Score

(%)

RL Model

F1-Score (%)

Conventional

Model

Accuracy

Time (ms)

1 tMRI_001 95.2 94.4 93.8 95.3 94.5 78.8 131.2

2 tMRI_002 95.2 95.9 92.4 90.1 93.6 83.6 129.8

3 tMRI_003 95.2 83.5 87.8 92.6 90.5 84.3 147.3

4 tMRI_004 94.6 91.8 91.2 82.7 92.1 84.1 132.5

5 tMRI_005 93.4 89.1 88.8 90.3 91.0 82.6 134.4

6 tMRI_006 94.5 88.8 90.2 89.4 92.2 83.8 135.3

7 tMRI_007 94.3 86.6 89.7 92.1 91.9 81.7 137.4

8 tMRI_008 94.1 85.5 88.3 93.2 91.1 85.1 133.2

9 tMRI_009 93.8 86.8 91.4 90.8 91.5 85.2 142.3

10 tMRI_010 94.6 85.5 90.3 93.4 91.8 87.6 144.5

11 tMRI_011 94.6 83.8 92.5 89.5 93.5 81.9 139.8

12 tMRI_012 94.6 84.8 92.3 90.3 90.2 83.2 146.3

ICRDICCT‘25 2025 - INTERNATIONAL CONFERENCE ON RESEARCH AND DEVELOPMENT IN INFORMATION,

COMMUNICATION, AND COMPUTING TECHNOLOGIES

692

Table 1 The RL models performed the best, with

precision of 87.8%-95.2%, recall of 85.9%-94.3%,

and F1 scores of 87.0%-93.7%, while detection times

were 12.1 ms-17.3 ms; the Conventional models

achieved precision of 82.5%- 88.9%, recall of 75.1%-

82.7% and F1 scores of 78.5%- 87.5% with the

detection times 135.5 ms-151.2 ms.

Table 2: Descriptive Statistics of Accuracy Time for RL Model and Conventional Model.

Model N Mean Std. Deviation Std. Error Mean

Accuracy Time RL Model 12 144.67 1.97 0.57

Conventional 12 145.10 4.73 1.36

Table 2 The RL models had a mean accuracy time

in: of 14.67 ms; std: 1.97; std. error mean: 0.57. On

the contrary, the Conventional models had more

mean accuracy times, with a mean of 145.10 ms,

standard deviation of 4.73 and standard error mean of

1.36. Independent sample test T-test Comparison of

the accuracy in RL Model and Conventional Model

Shown in Table 3.

Table 3: Independent Sample Test t-Test Comparison of the Accuracy in RL Model and Conventional Model.

Levene's Test for

Equality of Variances

Independent Samples Test

Accuracy Time F Sig. t df

Sig. (2-

tailed)

Mean

Difference

Std. Error

Difference

95% CI

Lowe

95% CI

Uppe

Equal variances

assume

0.627 0.438 48.197 22 1.000 130.425 3.848 127.5366 133.3234

Equal variances

not assume

48.197 11 1.000 130.425 3.848 124.9467 135.2634

Figure 2: Precision Comparison Between RL and

Conventional Models.

Figure 2 The Precision of the RL Model and the

Conventional Model over multiple iterations. The RL

Model demonstrates higher precision of 94%

compared to the Conventional Model's 86%.

Figure 3: Recall Comparison Between RL and

Conventional Models.

Figure 3 The comparison of recall of the RL

Model and the Conventional Model over multiple

iterations. The RL Model demonstrates higher recall

of 93% compared to the Conventional Model's 82%.

A Novel Method to Improve the Prediction of Vehicle Numbers Involved in Crashes at Rural Areas Using Reinforcement Learning Models

693

Figure 4: F1 Score Comparison Between RL and

Conventional Models.

Figure 4 The comparison of the F1-Score of the

RL Model and the Conventional Model over multiple

iterations. The RL Model demonstrates a higher F1-

Score of 94% compared to the Conventional Model's

86%.

Figure 5: Accuracy Time Comparison Between RL and

Conventional Models.

Figure 5 The compression of the detection time of

the RL Model and the Conventional Model over

multiple iterations. The RL Model demonstrates a

lower detection time of 15 ms compared to the

Conventional Model's 145 ms.

Figure 6: Accuracy Time Comparison.

Figure 6 The comparison of the accuracy time for

RL Model and Conventional Model. And here again,

the Conventional Model has higher accuracy time,

150, and the RL Model has significantly lower,

around 10.

6 DISCUSSION

The reinforcement learning (RL) based prediction

system for the number of vehicles involved in rural

area related crashes ensured the reliability and

improved performance of the predictions over other

similar models. Acquiring them helps a lot in a world

where, ideally, the goal should be prediction accuracy

between 87.8% and 95.2% with a processing time of

between 12.1ms and 17.3ms. Zhang G., et al., 2024

By utilizing these technologies, the systems can

perform in real-time and can play a crucial role in

enhancing road safety in remote locations through

timely and accurate predictions. More sophisticated

methods comprise RL algorithms that can work

alongside real-time traffic and other data to

dynamically modify predictions based on the current

scenario. Anand Kumar G., et al., 2025 More

sophisticated data fusion methods, such as integrating

traffic data to weather data, can increase the accuracy

of predictions across different environments.

Zhang C., et al., 2025 An IoT-enabled system for

real-time data collection and emergency notifications

may increase the responsiveness of this notification

system by enabling the exchange of useful data and

updates while improving the safety of users in such

situations. Karanikas N, et al., 2020 The use of

Reinforcement learning for real time data processing

ICRDICCT‘25 2025 - INTERNATIONAL CONFERENCE ON RESEARCH AND DEVELOPMENT IN INFORMATION,

COMMUNICATION, AND COMPUTING TECHNOLOGIES

694

properly ensures correctness, improving the accuracy

of predictions and providing timely alerts, is the main

goal of the project implemented using the

reinforcement learning models Reddy JS, et al., 2025.

Vinoth B., et al, 2025. Feedback mechanisms such as

the real time alerts used could lead to surplus

information for some users, Pusuluri VL, Dangeti

MR., 2024 while most existing models were

developed with an analysis of static data in mind and

are hence poorly equipped to deal with dynamism.

Zhang C, 2025., With timely alerts the authorities can

make informed and timely decisions to prevent

crashes in both urban and rural areas. It optimizes

prediction accuracy by continuously processing data

and making real-time changes. Going forward, the

system would be optimized by combining

lightweight, energy efficient hardware with

sophisticated RL algorithms to make the system more

feasible, while using adaptive AI models for real-time

data to detect recurring patterns and deviations more

accurately.

7 CONCLUSIONS

The prediction system based on reinforcement

learning (RL) shows effective performance and

improvement in predicting the number of vehicles

involved in roadside crashes of the rural areas. The

RL model proves better than older statistical models

in both precision and reliability while also ensuring

higher accuracy (87.8% to 95.2%) and lower

processing time (12.1ms to 17.3ms), which can

potentially lead to higher road safety in rural areas.

By utilizing RL in combination with real-time traffic

and environmental data, predictions can be modified

and improved over time; hence, predictions can be

made more accurately and timely. The approach is a

major leap forward in crash prediction and

management, particularly for rural environments,

where many trails and local betterment projects risk

crashes with motorized vehicle traffic.

REFERENCES

A literature review of machine learning algorithms for crash

injury severity prediction. Journal of Safety Research.

2022;80: 254–269.

A study on road accident prediction and contributing factors

using explainable machine learning models: analysis

and performance. Transportation Research

Interdisciplinary Perspectives. 2023;19: 100814.

Anand Kumar G, Mohiddin MK, Mishra SK, Verma A,

Sharma M, Naresh A. Enhancing Autonomous Vehicle

Navigation in Complex Environment with Semantic

Proto-Reinforcement Learning. Journal of Field

Robotics. 2025 [cited 27 Feb 2025].

doi:10.1002/rob.22506

Dhinesh Kumar R, Rammohan A, Sherazi HHR, Khan ZA.

Optimizing Intersection Safety through Next-Gen

Vehicular Communications: A Simulation-Based

Evaluation of Intersection Movement Assist Systems.

[cited 27 Feb 2025].

Available: https://ieeexplore.ieee.org/abstract/docume

nt/10574632

Jaradat S, Elhenawy M, Paz A, Alhadidi TI, Ashqar HI,

Nayak R. A Cross-Cultural Crash Pattern Analysis in

the United States and Jordan Using BERT and SHAP.

Electronics. 2025;14: 272.

Karanikas N, Chatzimichailidou M. Safety Insights:

Success and Failure Stories of Practitioners. CRC

Press; 2020.

Khan SS, Rahman MS, Rupak AUH, Rahman MS.

DeepInsureAI: A Deep Learning-Based Vehicle

Insurance Prediction Model. Innovations in Electrical

and Electronics Engineering. 2025; 273–285.

Machine learning for predictions of road traffic accidents

and spatial network analysis for safe routing on

accident and congestion-prone road networks. Results

in Engineering. 2024;23:

102737. Website. Available: https://www.researchgat

e.net/publication/260114095_Vehicular_traffic_noise_

modeling_using_artificial_neural_network_approach

Poonia RC, Singh V, Nayak SR. Deep Learning for

Sustainable Agriculture. Academic Press; 2022.

Pusuluri VL, Dangeti MR. Applications of QGIS and

machine learning for road crash spot identification.

Earth Science Informatics. 2024;17: 2331–2346.

Qawasmeh BS. Safety Assessment for Vulnerable Road

Users Using Automated Data Extraction with Machine

Learning Techniques. Western Michigan University.

2024. Available:

https://scholarworks.wmich.edu/dissertations/4112

Reddy JS, Rashmi MR, Pin LH. Road Accident Prediction

in Highways using Machine Learning Algorithms.

[cited 27 Feb 2025]. Available:

https://ieeexplore.ieee.org/abstract/document/1072216

Shamim Kaiser M, Ray K, Bandyopadhyay A, Jacob K,

Long KS. Proceedings of the Third International

Conference on Trends in Computational and Cognitive

Engineering: TCCE 2021. Springer Nature; 2022. IEEE

Xplore Full-Text PDF: [cited 27 Feb 2025]. Available:

https://ieeexplore.ieee.org/stamp/stamp.jsp?arnumber=

1 0444919

The statistical analysis of crash-frequency data: A review

and assessment of methodological alternatives.

Transportation Research Part A: Policy and Practice.

2010;44: 291–305.

Vinoth B, Prakash VS, Shivakumar BN. Road Traffic

Accident Prediction in India Using Machine Learning

Algorithm Techniques. [cited 27 Feb

A Novel Method to Improve the Prediction of Vehicle Numbers Involved in Crashes at Rural Areas Using Reinforcement Learning Models

695

2025]. Available: https://ieeexplore.ieee.org/abstract/

document/10593917.

Zhang G, Zhou S, University of Alaska Fairbanks. Center

for Safety Equity, University of Hawaii at Manoa.

Department of Civil and Environmental Engineering.

Enhancing Vehicle Sensing for Traffic Safety and

Mobility Performance Improvements Using Roadside

LiDAR Sensor Data. University of Alaska Fairbanks.

Center for Safety Equity in Transportation (CSET);

2024 Jun. Report No.: INE/CSET 24.06.

Zhang C, Jindal A. RoadSafe: A Machine Learning Based-

Road Accident Prevention System. [cited

27 Feb 2025]. Available: https://ieeexplore.ieee.org/ab

stract/document/10622142

ICRDICCT‘25 2025 - INTERNATIONAL CONFERENCE ON RESEARCH AND DEVELOPMENT IN INFORMATION,

COMMUNICATION, AND COMPUTING TECHNOLOGIES

696