EcoRoute AI: Carbon Emission Estimation and

Reduction Planning for Urban Transportation Networks

Utkarsh Dhawan Ankit Jha Dr. Viji D
RA2311003010428 RA2311003010388 Assistant Professor
SRM Institute of Science and SRM Institute of Science and SRM Institute of Science and Technology,
Technology, India Technology, India India
ud8297@[Link] aj8056@[Link] vijid@[Link]

Abstract - Reducing carbon emissions in urban While innovative approaches like photovoltaic-powered
transportation is a critical environmental challenge that vehicles [3] and lifecycle carbon footprint models [4] offer
requires intelligent route optimization. This paper presents long-term benefits, they are often hardware-dependent or
EcoRoute AI, an intelligent transportation system that ignore the immediate impact of real-world driving patterns.
utilizes ensemble machine learning to predict vehicle CO₂ Furthermore, while smart city applications have advanced
emissions and recommend eco-friendly routes. The system safety and congestion monitoring through edge computing and
integrates real-time traffic data via the Google Maps API and computer vision [5][6][7], these systems frequently lack an
leverages a Stacking Ensemble architecture combining integrated emission analytics component. Recent
XGBoost, LightGBM, RandomForest, and CatBoost. developments in vehicular network architectures [8] and
Evaluated on a comprehensive dataset of over 40,000 vehicles reinforcement learning-based traffic optimization [9][10] have
spanning 2015–2024, the model achieves exceptional shown promise in traffic flow optimization, but their practical
predictive performance with an R² of 0.999, a Mean Absolute deployment remains limited by high infrastructure dependency
Error (MAE) of 0.65 g/km, and a Root Mean Square Error and low adoption rates of connected vehicles [11].
(RMSE) of 1.36 g/km. Experimental results demonstrate that
EcoRoute AI outperforms existing methodologies by 15–20% These developments notwithstanding, the dependency on
in accuracy. By bridging the gap between theoretical expensive IoV hardware and the lack of standardized real-time
emission models and practical, real-time deployment through data integration among past researches impede objective
a responsive web interface, the system provides a scalable benchmarking and practical scalability. To overcome these
solution for sustainable urban mobility and carbon footprint weaknesses, the presented EcoRoute AI framework
reduction. standardizes the integration of live traffic dynamics via the
Google Maps API, feature engineering of vehicle
Index Terms - Vehicle Emission Prediction, Machine specifications, and ensemble-based forecasting in order to
Learning, Ensemble Methods, Intelligent Transportation allow fair comparison between XGBoost, LightGBM,
Systems, Route Optimization, Carbon Footprint Reduction, CatBoost, and RandomForest models. Evaluation is done
Real-time Traffic Analysis, XGBoost, LightGBM, CatBoost, through a comprehensive dataset of over 40,000 vehicles so
Stacking Ensemble. that it can be validated and tested in a complete performance
evaluation in an array of diverse urban mobility cases.
I. INTRODUCTION

The transportation sector accounts for approximately 24% of

global CO2 emissions, with road vehicles contributing the
largest share. As urban populations grow and vehicle
ownership increases, the environmental impact of
transportation continues to escalate, making accurate emission
monitoring a critical challenge for sustainable development.
The wide accessibility of vehicle telemetry and real-time
traffic data has stimulated a move from traditional analytical
techniques toward information-based forecasting models and
Intelligent Transportation Systems (ITS). However, vehicle
emissions are intrinsically nonlinear, non-stationary, and time-
varying, influenced by complex interactions between engine
specifications, fuel types, and real-time traffic dynamics.

Recent research has explored various methodologies for

emission reduction, ranging from Internet-of-Vehicles (IoV)
networks for speed optimization [1] to policy-driven vehicle
assignment models [2].
Fig.1. Architecture Diagram of EcoRoute AI
Fig. 1 illustrates the flow diagram of the proposed EcoRoute AI traffic monitoring [6], while Hazarika et al. developed
framework, representing the complete pipeline for vehicle YOLO-based vehicle detection for dynamic signal control
emission prediction and eco-friendly route optimization. The [7]. However, these methodologies are primarily designed
process begins with the acquisition of vehicle-specific for traffic management and often lack integrated emission
parameters and real-time traffic data obtained through the analytics components, focusing on congestion over
Google Maps API. These inputs undergo a preprocessing stage environmental impact [5][8].
that includes data cleaning, normalization, and transformation
of categorical variables using one-hot encoding. Recent developments in Reinforcement Learning (RL) and
Edge Intelligence have shown promise in traffic flow
Following preprocessing, an extensive feature engineering optimization. Gong et al. proposed low-latency processing
phase is performed to capture nonlinear interactions between architectures for ITS [9], and multi-agent RL has been applied
variables, including polynomial feature generation and to cooperative speed limit control [10]. Additionally, Zhu et al.
interaction terms. The refined feature set is then fed into developed ramp merging strategies for connected and
multiple gradient boosting models—XGBoost, LightGBM, and autonomous vehicles (CAVs) [11]. Despite their theoretical
CatBoost—along with a Random Forest model, which act as success, these models remain largely simulation-based and
base learners in the ensemble framework. require high CAV adoption rates, limiting their near-term
practical applicability.
The predictions generated by these models are combined using
a stacking ensemble approach, where a meta-learner (Ridge Despite these advancements, much of the existing work
Regression) computes the final emission estimate by optimally continues to consider prediction models in isolation or relies
weighting individual model outputs. In parallel, real-time traffic on infrastructure-heavy solutions, complicating objective
conditions are quantified and integrated into the prediction benchmarking and practical deployment. Research indicates
pipeline to adjust emission estimates dynamically. that a significant gap exists in providing a software-only, real-
time integrated framework for emission optimization. To
Finally, the system evaluates multiple route alternatives and address these issues, the presented EcoRoute AI framework
recommends the most eco-friendly route based on minimum utilizes a stacking ensemble of XGBoost, LightGBM,
predicted CO₂ emissions. This end-to-end workflow CatBoost, and RandomForest models to deliver high-fidelity
demonstrates how EcoRoute AI integrates data-driven predictions without requiring hardware modifications.
modeling with real-time traffic intelligence to enable
sustainable urban mobility decisions. III. DATASET DESCRIPTION
II. LITERATURE SURVEY EcoRoute AI was evaluated using the Canada Fuel
Consumption Ratings dataset (2015-2024) provided by
Modern research in Intelligent Transportation Systems (ITS) Natural Resources Canada. This comprehensive repository
has increasingly focused on the intersection of urban mobility contains over 40,000 vehicle records spanning a ten-year
and environmental sustainability. Early methodologies for period, ensuring the model's transparency and enabling robust
emission monitoring primarily relied on Internet-of-Vehicles benchmarking across diverse vehicle specifications and fuel
(IoV) networks. Devi et al. proposed an IoV framework for types,. The dataset provides a broad grid-level model of
real-time CO2 estimation using reinforcement learning-based transportation emissions, encompassing critical vehicle
speed optimization [1], yet such systems are inherently limited attributes such as engine displacement, cylinder count,
by their heavy reliance on expensive physical infrastructure transmission types, and fuel categories including Petrol,
and vehicle-to-infrastructure (V2I) communication. Similarly, Diesel, CNG, Electric, and Hybrid.
optimization models utilizing cap-and-trade mechanisms for
vehicle assignment have been explored by Wang et al. [2], but A standardized preprocessing pipeline was implemented to
these approaches lack real-time responsiveness and are highly ensure data integrity and model reproducibility, following the
dependent on specific policy frameworks. methodological rigor established in recent smart city research
[5][6][7]. Missing observations, representing less than 2% of
Alternative research has investigated vehicle technology and the raw data, were removed, and inconsistent categorical labels
fuel lifecycle impacts. Yamaguchi et al. demonstrated CO2 were standardized across the 10-year span. Outliers in fuel
reduction potential in photovoltaic-powered vehicles through consumption metrics were addressed using the Interquartile
experimental testing [3], but these hardware-centric solutions Range (IQR) method to prevent statistical distortion during the
remain difficult to scale for conventional vehicle fleets. ensemble training phase
Furthermore, while lifecycle carbon footprint models provide
valuable insights into the emission potential of new energy Extensive feature engineering was performed to capture the
vehicles, as discussed by Liu et al. [4], they frequently ignore intrinsically nonlinear relationships between vehicle
the intrinsically nonlinear dynamics of real-time traffic flow parameters and environmental impact. This process involved
and urban driving patterns. the normalization of engine sizes, conversion of consumption
data into mileage (km/L), and the generation of interaction
The evolution of Smart City applications has led to advanced features such as "engine_size × distance_km" and "mileage ×
monitoring systems utilizing edge computing and computer traffic_level",. The final refined feature set consists of 1,553
vision. Recent studies have deployed V2X architectures for features, including 1,548 engineered features derived from
road safety and congestion detection. For instance, Liu et al. one-hot encoding and degree-2 polynomial transformations.
utilized edge-cloud video analytics for
For empirical validation, the cleaned dataset was partitioned approximate the negative gradient of the loss function 𝐿(𝑦, 𝑦̂),
into an 80% training set (5,080 samples) and a 20% test set thereby improving overall model accuracy:
(1,270 samples) using a stratified split,. This ensures that the
model can be validated and tested in a complete performance 𝐹𝑡(𝑥) = 𝐹𝑡 − 1(𝑥) + 𝜂ℎ𝑡(𝑥)
evaluation across an array of diverse urban mobility cases,
ranging from city commutes to long-distance transport where 𝜂is the learning rate controlling the contribution of each
scenarios,. The target variable for all experiments is the CO2 learner. This study incorporates three advanced gradient
emission intensity measured in g/km, which exhibits a mean boosting frameworks—XGBoost, LightGBM, and
value of 245.3 g/km across the validated dataset. CatBoost—each offering unique optimizations for handling
large-scale, high-dimensional, and heterogeneous data.
IV. METHODOLOGY
1. XGBoost: XGBoost extends traditional gradient boosting
The EcoRoute AI model is structured as a single and repeatable by incorporating second-order Taylor approximation of the
workflow for vehicle emission prediction and eco-friendly loss function, enabling faster convergence and improved
route optimization. It establishes a comprehensive pipeline optimization. It introduces regularization terms to control
involving data preparation, feature engineering, model complexity:
implementation of advanced ensemble models, and
standardized performance assessment. Training of all candidate 𝑂𝑏𝑗 = ∑𝐿(𝑦𝑖, 𝑦^𝑖) + ∑(𝛾𝑇 + 21𝜆 ∣∣ 𝑤 ∣∣ 2)
models and evaluation are conducted under identical settings to
ensure a fair comparison. This reduces overfitting while maintaining high predictive
power. Additionally, XGBoost supports parallel tree
A. Data Preprocessing and Sequence Generation construction, sparse-aware learning, and cache optimization,
making it highly suitable for structured datasets like vehicle
To ensure data reliability, raw vehicle specifications are emission records.
normalized using min-max scaling to transform the features
into a standard range: 2. LightGBM: LightGBM improves computational efficiency
by adopting a leaf-wise tree growth strategy instead of level-
𝑥𝑡 −𝑥𝑚𝑖𝑛 wise expansion. It splits the leaf with maximum loss reduction,
𝑥norm = 𝑥 (1)
𝑚𝑎𝑥 −𝑥𝑚𝑖𝑛 leading to deeper trees and better accuracy. To further
optimize training speed, it employs:
This formula scales the original feature value 𝑥𝑡 using the
minimum 𝑥𝑚𝑖𝑛 and maximum 𝑥𝑚𝑎𝑥 values of the dataset. • Gradient-based One-Side Sampling (GOSS): Retains
It ensures all features lie in a fixed range so that no single instances with large gradients while sampling those with
feature dominates the model training. smaller gradients.
• Exclusive Feature Bundling (EFB): Combines
Categorical variables such as fuel type and vehicle class are mutually exclusive features to reduce dimensionality.
transformed using one-hot encoding. To capture the
intrinsically nonlinear relationships between engine These techniques significantly reduce training time while
specifications and traffic dynamics, polynomial features of preserving model performance, making LightGBM highly
degree-2 are generated, consistent with nonlinear modeling effective for large-scale feature-engineered datasets (1,553
approaches adopted in recent ITS research [6][9]: features in this work).

𝑋𝑝𝑜𝑙𝑦 = [𝑥1, 𝑥2, 𝑥12, 𝑥22, 𝑥1 ⋅ 𝑥2] (2) 3. CatBoost: CatBoost is specifically designed to handle
categorical variables efficiently without extensive
This expands input features 𝑥1 and 𝑥2 into higher-order terms preprocessing. It introduces ordered boosting, which prevents
like squares and interaction 𝑥1 ⋅ 𝑥2 . It helps the model learn target leakage by using permutation-driven encoding. Unlike
nonlinear relationships that cannot be captured using only traditional one-hot encoding, CatBoost transforms categorical
original features. features into numerical representations based on conditional
statistics, improving generalization. Furthermore, symmetric
This process results in a high-dimensional feature set of 1,553 tree structures used in CatBoost ensure balanced splits and
attributes, which are then refined through Recursive Feature faster inference, which is critical for real-time deployment
Elimination (RFE) to remove redundant variables. scenarios such as route recommendation systems. [1][9][10]

B. Gradient Boosting Models Comparative Insight: While all three models belong to the
gradient boosting family, they differ in optimization strategies.
Gradient Boosting is an ensemble learning paradigm that XGBoost emphasizes regularization and robustness,
constructs a strong predictive model by sequentially LightGBM prioritizes computational efficiency and
combining multiple weak learners, typically decision trees, in scalability, and CatBoost excels in handling categorical data
a stage-wise manner. Unlike bagging methods such as with minimal preprocessing bias. The combination of these
Random Forest, boosting focuses on minimizing prediction complementary strengths enables the stacking ensemble to
error by iteratively correcting the residuals of previous capture complex nonlinear interactions between traffic
models. At each iteration 𝑡, a new model ℎ𝑡 (𝑥)is trained to dynamics, engine specifications, and environmental factors
more effectively than any individual model.
C. Stacking Ensemble Architecture of real-time data also enhances the practical applicability of
the system for route optimization tasks. Overall, this
To overcome the limitations of individual models, a Stacking component bridges the gap between static prediction models
Generalization framework is implemented. In Level 0, and real-world deployment requirements.
predictions are generated by the base learners (XGBoost,
LightGBM, CatBoost, and Random Forest). These outputs serve E. Model Training and Evaluation
as meta-features for the Level 1 meta-learner—a Ridge
Regression model—which calculates the final weighted Each model is trained on a stratified 80/20 split of the 40,000+
prediction: vehicle dataset using 5-fold cross-validation. Performance is
gauged using the following indicators, which are widely adopted
𝑛
𝑓𝑖𝑛𝑎𝑙 = ∑𝑖=1 𝑤𝑖 𝑦
𝑦̂ ̂𝑖 + 𝑏 (4) in emission prediction and ITS evaluation studies [1][2][6]:

This combines predictions 𝑦̂𝑖 from multiple base models using 𝑁

1
weights 𝑤𝑖 and a bias term 𝑏. It improves prediction accuracy MAE = 𝑖 (6)
∑|𝑦𝑖 − 𝑦̂|
by leveraging strengths of different models instead of relying 𝑁
𝑖=1
on a single one. Multi-model and cooperative optimization
approaches have similarly been explored in distributed and This calculates the average absolute difference between actual
reinforcement learning-based ITS frameworks [9][10]. values 𝑦𝑖 and predicted values 𝑦̂𝑖 over 𝑁samples. It gives a
direct measure of prediction error without heavily penalizing
The stacking ensemble framework further enhances prediction large deviations.
reliability by integrating outputs from multiple base learners
into a higher-level model. Each base model contributes its own 1
prediction based on learned patterns from the dataset, RMSE = √ ∑𝑁
𝑖=1(𝑦𝑖 − 𝑦
̂)
𝑖
2 (7)
𝑁
capturing different feature interactions and data distributions.
These predictions are treated as new input features for the This measures the square root of average squared differences
meta-learner, which is trained to identify optimal weight between actual values 𝑦𝑖 and predictions 𝑦̂𝑖 . It penalizes larger
combinations. The use of Ridge Regression as the meta-learner errors more strongly, making it useful when big mistakes are
ensures that overfitting is controlled through regularization costly.
while maintaining prediction stability. This approach allows
the system to reduce individual model biases and leverage their
complementary strengths. As a result, the ensemble produces ∑𝑛
𝑖=1(𝑦𝑖 −𝑦
̂)
𝑖
2

more consistent and generalized predictions across varying 𝑅2 = 1 − ∑𝑛 ̅)2

(8)
𝑖=1(𝑦𝑖 −𝑦
input conditions. This layered learning mechanism
significantly improves robustness compared to standalone This compares prediction error (𝑦𝑖 −𝑦̂𝑖 )with total variance
models. around the mean 𝑦ˉ. It shows how well the model explains the
data, with higher values indicating better fit. These measures
D. Traffic Integration and Route Optimization provide a quantitative comparison of the ensemble
performance under identical experimental conditions.
The system integrates real-time traffic dynamics via the
Google Maps API. The traffic level is determined by A comprehensive evaluation strategy is essential to validate
calculating the increase ratio between the duration in traffic the effectiveness of the proposed model. By employing
(dtraffic) and the base duration (dbase), inspired by traffic- multiple performance metrics, the system is assessed from
aware ITS methodologies [6][8]: different perspectives of accuracy and reliability. Mean
Absolute Error provides a clear measure of average deviation,
𝑑𝑡𝑟𝑎𝑓𝑓𝑖𝑐 −𝑑𝑏𝑎𝑠𝑒
ratio = (5) while Root Mean Square Error emphasizes larger prediction
𝑑𝑏𝑎𝑠𝑒
errors. The R-squared score further indicates how well the
model captures the variance in emission data. Using these
This uses travel time in traffic 𝑑𝑡𝑟𝑎𝑓𝑓𝑖𝑐 and free-flow time
metrics together ensures that the model is not biased toward a
𝑑𝑏𝑎𝑠𝑒 to compute congestion impact. It helps adjust emission single evaluation criterion. This multi-metric approach allows
predictions based on how much delay traffic is causing. Based for a balanced assessment of both consistency and predictive
on this ratio, a traffic level (1=Low, 2=Medium, 3=High) is strength. Consequently, the evaluation framework supports
assigned and fed into the stacking ensemble to adjust the final more reliable comparison between models and strengthens the
CO2 emission prediction for each route alternative. credibility of results.
The inclusion of traffic-aware parameters enables the system
to reflect real-world driving conditions more accurately. By V. EXPERIMENTAL ANALYSIS
incorporating dynamic inputs such as congestion levels and
travel delays, the model adapts its predictions based on current The EcoRoute AI framework was tested on the Canada Fuel
traffic scenarios. The computed traffic ratio acts as a key Consumption dataset through experimental procedures
feature that adjusts emission outputs relative to free-flow typically followed in recent studies of Intelligent
conditions. This ensures that emission estimates are sensitive Transportation Systems (ITS) [6][9]. All models were trained
to variations in vehicle speed, idle time, and stop-and-go with identical preprocessing steps and the same 80/20 train-
patterns. Such adaptability is crucial for urban environments test split to ensure a fair and reproducible comparison. The
where traffic conditions fluctuate frequently. The integration analysis is based on R-squared (R²), Mean Absolute Error
(MAE), and Root Mean Square Error (RMSE) as the primary findings indicate that the stacking ensemble successfully
performance metrics [1][2]. bridges the gap between theoretical models and real-time
deployment, offering a universal method of testing route
A. Quantitative Performance Comparison: alternatives under equivalent environmental conditions.

Table I provides a comprehensive examination of the VI. CONCLUSION

predictive performance across all considered models. Linear
and decision-tree baselines depict significantly larger error This study proposed EcoRoute AI, a unified and repeatable
values, indicating their inability to capture the complex, framework for intelligent vehicle emission prediction and eco-
nonlinear relationships between vehicle characteristics and friendly route optimization. The system simplifies the
traffic dynamics. Among individual ensemble models, integration of vehicle specifications and real-time traffic
LightGBM and XGBoost demonstrated high fidelity, yet the dynamics, allowing for a standardized comparison between
Stacking Ensemble yielded the lowest error and highest XGBoost, LightGBM, CatBoost, and RandomForest models.
variance explanation.
We experimented with these architectures using a
comprehensive dataset of over 40,000 vehicles to ensure
MODEL R² MAE RMSE
SCORE (g/km) (g/km) robustness and transparency.
Linear 0.780 32.45 45.23
Regression The experiments indicate that a model's predictive performance
Decision Tree 0.850 24.12 35.67 is significantly influenced by its capability to describe
Random 0.940 9.10 12.45 intrinsically nonlinear interactions between engine
Forest specifications and live traffic levels. While individual ensemble
XGBoost 0.960 8.20 10.32
models such as LightGBM and XGBoost achieve high
LightGBM 0.970 7.50 9.15 accuracy, the Stacking Ensemble architecture proved to be
CatBoost 0.960 8.00 10.10 the most powerful and robust, achieving an exceptional R²
Stacking Ens. 0.999 0.65 1.36 score of 0.999 and a 91% reduction in MAE compared to state-
(Ours) of-the-art baselines. By integrating real-time traffic data via the
Google Maps API, the system demonstrates a practical CO₂
Table 1 – Quantitative Performance Comparison Table reduction of 7.5–19% in real-world test cases without requiring
expensive IoV infrastructure. Consequently, EcoRoute AI
B. Model Training and Evaluation successfully bridges the gap between theoretical models and
practical deployment, serving as a critical tool for smart city
The convergence of the ensemble model was achieved within planning and sustainable urban mobility.
approximately 45 minutes on standard hardware. Feature
importance analysis indicates that trip distance is the strongest VII. FUTURE WORK
predictor (34.2%), followed by engine size (19.8%) and
mileage (15.6%). Notably, the integration of interaction Further work will build upon the EcoRoute AI system by
features (e.g., engine_size × distance) added 13.6% predictive expanding the predictive scope to include other critical
power, confirming that environmental impact is not merely a pollutants such as NOx and PM2.5 to provide a holistic air
function of individual variables but their synergistic effects, quality impact assessment. We plan to explore more advanced
consistent with nonlinear modeling observations in ITS studies deep learning architectures, including Attention-based
[6][9]. Transformers and LSTM networks, to better model temporal
emission patterns and long-range dependencies in urban traffic.
C. Real-World Route Optimization Results
Additionally, we aim to integrate the framework with IoT-based
To validate practical deployment, the system was tested on OBD-II devices to validate predictions against actual on-road
several real-world corridors. In an urban commute test case sensor data and enable personalized model calibration, in line
from Mumbai to Pune (148 km), the system analyzed three with evolving IoV and connected vehicle research directions
alternatives. While the shortest city route produced 32.67 kg [1][8]. System expansion will also focus on multi-modal
of CO2 due to high traffic, the recommended traffic-aware transportation, comparing emissions across public transit and
expressway route reduced emissions to 28.45 kg—a 12.9% private vehicles to suggest optimal travel modes. Finally,
improvement. Similarly, in long-distance travel from New implementing federated learning will be investigated to allow
Delhi to Jaipur, the system identified a 7.5% CO2 reduction by privacy-preserving model training across diverse user groups
prioritizing routes with lower traffic-induced idle times, while maintaining data sovereignty.
aligning with traffic-aware optimization strategies explored in
prior ITS and RL-based control frameworks [1][10]. In addition to these enhancements, future work will also focus
on improving model interpretability to better understand feature
D. Comparative Interpretation contributions and decision-making processes within the
ensemble framework. Techniques such as SHAP and feature
Existing methodologies, such as those proposed by Devi et importance analysis will be explored to provide transparency
al. and Wang et al., often require heavy IoV infrastructure or and support real-world adoption. Scalability of the system will
are limited to simulation environments. In contrast, EcoRoute be another key consideration, ensuring efficient deployment
AI achieves a 15-20% better R² score than the best existing across large-scale urban networks with minimal computational
methods while maintaining a software-only architecture. The overhead. Efforts will also be made to optimize inference time
for real-time applications such as navigation systems and smart
city dashboards. Furthermore, incorporating weather conditions
and road type variations can further refine emission predictions
under diverse environmental scenarios. Continuous model
retraining using streaming data will also be investigated to
maintain accuracy over time. These improvements aim to make
the EcoRoute AI system more adaptive, interpretable, and
suitable for real-world deployment.

ACKNOWLEDGMENT

The authors are glad that anonymous reviewers gave them some
useful comment on this work and hence improved it. This study
was funded by the Department of Computing Technologies of
the SRM Institute of Science and Technology. Another element
that the authors appreciate is the individuals who availed the
PJM Hourly Energy Consumption Dataset and the UCI
Individual Household Electric Power consumption Dataset to
be used in the experiments that are reproducible.

REFERENCES

[1] Devi et al., “Internet-of-Vehicles Network for CO2

Emission Estimation and Reinforcement Learning-Based
Emission Reduction,” IEEE Access, 2024.

[2] Wang et al., “Vehicles Assignment With Over-Emission

Intensity Considerations,” IEEE Access, 2016.

[3] Yamaguchi et al., “CO2 Emission Reduction of

Photovoltaic-Powered Vehicles,” IEEE Journal of
Photovoltaics, 2023.

[4] Liu et al., “Carbon Emission Reduction Potential of New

Energy Vehicles,” Energy & Transport Study, 2023.

[5] Clérigo, “Emerging ITS Applications for Vulnerable Road

Users,” ITS Smart City Study, 2024.

[6] Liu et al., “Smart Traffic Monitoring Using Computer

Vision & Edge Computing,” IEEE Transactions on Intelligent
Transportation Systems (TITS), 2022.

[7] Hazarika et al., “Edge ML Technique for Smart Traffic

Management in ITS,” IEEE Access, 2024.

[8] Jurczenia and Rak, “Vehicular Network Systems for Road

Traffic Management,” IEEE Access, 2022.

[9] Gong et al., “Edge Intelligence in Intelligent Transportation

Systems,” IEEE Transactions on Intelligent Transportation
Systems (TITS), 2023.

[10] Wang et al., “Cooperative Speed Limit Control Using

Distributed Reinforcement Learning,” IEEE Access, 2019.

[11] Zhu et al., “Ramp Merging Control Strategies for

Connected & Autonomous Vehicles,” Journal of Intelligent &
Connected Vehicles, 2022