Content uploaded by Apurva Pamidimukkala
Author content
All content in this area was uploaded by Apurva Pamidimukkala on Mar 14, 2024
Content may be subject to copyright.
1
Enhancing AV Safety: A Bagging Classifier Approach
for Predicting Crash Outcomes
Sai Sneha Channamallu,1 Deema Almaskati,2 Sharareh Kermanshachi, Ph.D., P.E.,3 and
Apurva Pamidimukkala, Ph.D.4
1Ph.D. student, Department of Civil Engineering, The University of Texas at Arlington, Arlington,
TX-76019. E-mail: sxc5928@mavs.uta.edu
2Ph.D. student, Department of Civil Engineering, The University of Texas at Arlington, Arlington,
TX-76019. E-mail: deemanabeel.almaskati@mavs.uta.edu
3Associate Vice Chancellor for Research and Associate Dean of Research, Pennsylvania State
University, State College, PA-16801. E-mail: svk5464@psu.edu
4Assistant Professor of Research, Department of Civil Engineering, The University of Texas at
Arlington, Arlington, TX-76019. E-mail: apurva.pamidimukkala@mavs.uta.edu
ABSTRACT
Safety is the predominant concern in the advancement of automated vehicles (AVs); therefore,
extensive real-world testing is underway to ensure their secure operation. Despite the widespread
belief that they will decrease the frequency of accidents, there remains uncertainty about their
impact on the severity of crashes in which they are involved. The primary objective of this study
is to use the bagging classifier technique to predict the likelihood of injuries in accidents involving
AVs. This was accomplished by conducting an in-depth examination of a wide range of
independent variables and an analysis of injuries sustained in crash incidents involving AVs from
2014 to July 2023. The bagging classifier model showed notable effectiveness, achieving a
balanced accuracy of 0.59, along with high precision and recall values of 0.94 and 0.97,
respectively. These metrics indicate the model's strong capability for accurately identifying severe
crash outcomes and minimizing false positives. The precision-recall curve and a modified F1 score
of 2.39 further endorse the model's performance, particularly highlighting its efficiency in handling
the class imbalance present in the dataset. The validation and learning curves underscore the
model's optimal complexity, displaying its proficiency in identifying essential patterns without
succumbing to overfitting. Collectively, these metrics underscore the model's success in predicting
injury outcomes in AV crashes with a high level of accuracy. This study contributes to the literature
on AV safety by providing valuable insights for manufacturers and policymakers that will enable
them to develop effective safety features and strategies, thereby enhancing traffic safety.
KEYWORDS: Autonomous vehicles, Crash injury, Predictive modeling, Crash patterns, Machine
learning, Bagging classifier.
INTRODUCTION
Autonomous vehicles (AVs) have the potential to transform safety in the transportation sector by
minimizing human involvement, the primary factor contributing to vehicle accidents (Novat et al.,
2023; Channamallu et al., 2023). They are also expected to enhance mobility, decrease emissions
and energy consumption, and reduce congestion (Song et al, 2021; Houseal et al., 2022;
2
Pamidimukkala et al., 2023a). Autonomous driving systems are intended to support and assist
human drivers by managing and controlling driving tasks (Favarò et al., 2017; Etminani-
Ghasrodashti et al., 2022a). The Society of Automotive Engineers (SAE) outlines six levels of
autonomy (L0-L5) that assess how much support and assistance AVs can provide for driving tasks
(Das et al., 2020; Channamallu et al., 2023a; Khan et al., 2022a). Though it is generally accepted
that they hold great promise for reducing injuries and fatalities due to roadway accidents, more
testing is necessary to gain a better understanding of their capabilities and to pinpoint areas that
require technological and operational improvements prior to commercialization (Kutela et al.,
2022; Pamidimukkala et al., 2023b; Patel et al., 2022a). Automated driving systems are currently
being tested globally, including in California, where the state has mandated that all manufacturers
testing on public roads must submit reports on disengagements and collisions to the California
Department of Motor Vehicles (CA DMV) (Ye et al., 2021; Khan et al., 2022b; Etminani-
Ghasrodashti et al., 2022b).
Researchers globally have utilized a wide range of statistical models and descriptive analyses
to evaluate the CA DMV data. For instance, Novat et al. (2023) developed a Bayesian network,
using the Markov Chain Monte Carlo to establish a comparative analysis between collisions of
AVs and conventional vehicles, and Ding et al. (2023) evaluated the variables impacting crash
severity by utilizing a multinomial logit model. Other researchers investigated crash severity
factors, utilizing ordinal regression, binary logistic regression, and probit models. While the
findings of these studies are essential for a better understanding of the performance of AVs in a
mixed traffic environment, statistical models can often be characterized by imbalanced data. This
study attempts to address this issue by incorporating class weighting strategies. Sinha et al. (2021)
is one of the few who has developed a crash model utilizing a bagging classifier; however, they
only assessed datapoints from 2014 to 2019, whereas this paper provides a comprehensive
assessment of the data from 2014 to July 2023.
The primary objective of this study is to develop a bagging classifier crash model that can
predict the outcome of injuries resulting from collisions involving AV s. This paper will contribute
to the current literature on autonomous driving systems and CA DMV data by providing a
comprehensive analysis of AV crash predictions, using a bagging classifier model. The inferences
from the model will be beneficial to AV manufacturers for creating more adaptable and responsive
safety features, which may lessen the severity of crashes and lead to increased traffic safety in
general. The findings will also equip transportation professionals and policymakers with a
thorough understanding of AV collisions and help them develop measures and policies that will
reduce their severity.
LITERATURE REVIEW
In 2014, the CA DMV authorized manufacturers to test AVs on public roads to expedite the
development of autonomous driving systems (Ye et al., 2021; Khan et al., 2023a; Patel et al.,
2023a). Since then, more crashes and disengagements have been reported to the DMV due to the
increase of AV roadway testing, and researchers have analyzed the data to evaluate AV
performance.
Previous studies have employed a range of analytical techniques, including a hierarchal
Bayesian approach (Ren et al., 2022; Patel et al., 2022b), classification tree (Wang and Li, 2019;
Zhu and Meng, 2022; Etminani-Ghasrodashti et al., 2023; Channamallu et al., 2023b), and probit
model (Yuan et al., 2022; Khan et al., 2023b); others have attempted to determine the risk factors
3
connected to AV collisions. Leilabadi and Schmidt (2019) examined how environmental
conditions affect the performance of AVs, and the results showed that there is no significant
relationship between the severity of crashes and the roadway conditions or the weather (Leilabadi
and Schmidt, 2019; Khan et al., 2022c). The lack of correlation may not be representative of the
performance of AVs in a different climate, however, since California is not characterized by its
adverse weather conditions. Similarly, there was minimal association with lighting conditions, a
factor which Ye et al. (2021) considered to be the most significant in traffic injuries. The Pearson
Chi-Square test conducted by Leilabadi and Schmidt (2019) found roadway surfaces to have the
strongest correlation with AV performance. Houseal et al. (2022) examined AV collisions and,
through logistic regression and decision trees, identified the type of collision and movements of
AVs and other vehicles as significant contributing factors to AV-involved accidents. Similarly,
Wang and Li (2019) utilized crash database and statistical modeling approaches such as ordinal
logistic regression and CART classification tree to investigate the contributing factors and
mechanisms of AV crashes and found that crash severity significantly increased when the AV was
primarily responsible for the collision. Wang and Li (2019) and Patel et al (2023b) determined that
individuals are more likely to suffer severe injuries when accidents occur on highways, possibly
due to the high travel speeds. In contrast, Favarò et al. (2017) found that only 20% of accidents
occur on highways; most (48%) had taken place on suburban roads at low relative speeds. By
employing text analytics and a hierarchal Bayesian regression to analyze crash reports, Boggs et
al. (2020) concluded that AVs are more likely to be involved in a collision in areas of mixed land
use in comparison to other land uses, such as public areas and school zones.
METHODOLOGY
The methodology adopted in this study to develop a bagging classifier model to predict the
outcomes of injuries sustained in AV crashes is presented in Figure 1. Data collection and data
processing, including cleaning, feature elimination, and transformation of categorical data into
numerical formats were employed, and evaluation metrics were selected to provide a nuanced
assessment of the model's predictive performance in the context of skewed data distributions.
Figure 1. Methodology
Data Collection and Pre-processing
The study leverages the dataset originally compiled from the CA DMV, mandating that companies
testing AVs on public roads submit accident reports in PDF format within 10 business days. The
process included accumulating specifics from every PDF report on the CA DMV's website,
spanning from 2014 to July 2023. After extracting these details, they were methodically compiled
into an Excel file for a numerical examination. Figure 1 presents the yearly distribution of AV
crashes in the data set.
4
Figure 2. Yearly distribution of AV crashes
To ensure the relevance and quality of the data, it was meticulously cleaned and
preprocessed, including the removal of irrelevant features such as exact location coordinates that
do not directly contribute to crash outcomes. Missing values were addressed by excluding records
with significant gaps, and categorical variables were transformed into numerical formats using
techniques such as one-hot encoding, to facilitate computational analysis and model training.
Figure 3 presents a correlation matrix of the selected dependent variables.
Figure 3. Correlation matrix of variables
The feature selection for the model was informed by the relevance and contribution of
various factors to crash outcomes. These factors include the extent of vehicle damage, the AV
manufacturer, the nature of the collision, and environmental conditions like weather and lighting,
5
along with the time attributes of the crash. The correlation matrix presents a quantitative analysis
of how different variables relate to each other, ranging from technical aspects like vehicle type and
AV status, to environmental conditions such as weather and lighting. Variables such as the status
and mode of the AV have a high positive correlation (0.77), suggesting that the autonomous mode
of the vehicle is strongly related to its operational status. Hence, mode was excluded from the
variables.
Model Development
The crash prediction model that was developed included one dependent and 15 independent
variables including vehicle-related, environmental, traffic-related and contextual factors for
forecasting injury outcomes in crashes. The dependent variable was binary, coded as 0 for crashes
without injuries and 1 for those with injuries. There was a notable imbalance in the sample sizes
of the two categories, with 331 non-injury incidents compared to just 27 injury incidents. To
counter this imbalance, the model employed class weighting techniques, assigning weights to each
class inversely related to their occurrence in the dataset. The weighting formula shown in Equation
1 factors in the total number of samples (n_sample), the number of classes in the dataset
(n_classes), and the count of each class (np.bincount(y)).
Class weights of a specific class y = 𝑛_𝑠𝑎𝑚𝑝𝑙𝑒
𝑛_𝑐𝑙𝑎𝑠𝑠𝑒𝑠 𝑥 𝑛𝑝.𝑏𝑖𝑛𝑐𝑜𝑢𝑛𝑡(𝑦) (1)
The dataset was divided into training and testing portions, with a distribution of 80% for
training purposes and 20% for testing. Stratified sampling was applied to counteract the
imbalances in class distribution and maintain equal class proportions in both training and testing
sets. A bagging classifier was selected as the main predictive model due to its ability to handle
complex data and resist overfitting. During training, the model's parameters, including the number
of decision trees and their depth, were fine-tuned through grid search and cross-validation,
ensuring a balance between bias and variance.
Evaluation Metrics
The model's performance was assessed, using sophisticated metrics such as balanced accuracy that
better reflect its effectiveness in datasets with skewed distributions than traditional accuracy
measures. Precision and recall were deemed crucial metrics, given the importance of accurately
predicting crash severity, and they provided insights into the model’s proficiency in accurately
identifying true positive cases, while reducing false negatives and positives. The F1-Score, which
is the harmonic mean of precision and recall, was utilized to offer a comprehensive perspective on
the model's predictive capabilities. Minimizing false positives was a key focus of the model, given
the significant implications of incorrectly predicting non-injury incidents. Therefore, the study
adopted an adapted evaluation approach that uses the modified F1 score as the principal metric, as
recommended in prior research (Li et al., 2018; Sinha et al., 2021). This score effectively balances
precision and recall, and provides reliable predictions across both classes, a crucial aspect in
handling imbalanced datasets. The computation for the modified F1 score is detailed in Equation
2 where β is a parameter; When β is a parameter with a value between 0 and 1, the focus is more
on precision, whereas a β greater than 1 indicates that recall is more important. In this study, a β
of 0.5 is employed to place greater importance on precision.
6
Modified F1 score = (1+β2) x 𝑝𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛 𝑥 𝑟𝑒𝑐𝑎𝑙𝑙
β2 x (𝑝𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛+𝑟𝑒𝑐𝑎𝑙𝑙) (2)
DISCUSSION
The bagging classifier's performance in predicting AV crash outcomes is presented in Table 1. The
model achieved a balanced accuracy of 0.59, which is particularly significant as it indicates its
ability to maintain an equilibrium between sensitivity (true positive rate) and specificity (true
negative rate). Balanced accuracy is critical in AV crash predictions due to the dual importance of
correctly identifying both non-severe and severe outcomes.
Table 1 Results of bagging classifier model
Confusion matrix
Evaluation metrics
Balanced
accuracy
Precision
Recall
F1
score
Modified
F1 score
0.59
0.94
0.97
0.96
2.39
The precision score of 0.94 demonstrates the bagging classifier’s adeptness for accurately
classifying positive instances (severe crash outcomes) and is indicative of the model’s ability to
minimize false positives - a key factor in avoiding unnecessary alarms or interventions. The recall
score of 0.97 highlights the model’s proficiency in detecting actual positive cases and is essential
in AV safety contexts, where failing to identify a severe crash outcome could have critical
implications. The combination of high precision and recall in the bagging classifier is particularly
noteworthy, as it suggests that the model is not only good at identifying severe crash outcomes
when they occur but also in correctly recognizing non-severe outcomes, which minimizes the risk
of misclassification. The modified F1 score of 2.39, adjusted to account for class imbalance, stands
out as a commendable achievement. This adjustment is crucial because conventional F1 scores
may not fully capture the model's performance in datasets where one class (e.g., severe crashes) is
significantly underrepresented. This high modified F1 score indicates that the bagging classifier is
effective in balancing precision and recall, even in an imbalanced dataset. This is a significant
strength, as imbalanced data is a common challenge in AV crash datasets where severe outcomes
are, fortunately, less frequent.
Threshold Analysis
The precision-recall curve presented in figure 4 is an essential visualization tool that offers
profound insights into the predictive performance of our bagging classifier model, as it reflects
the classifier's efficacy in distinguishing between severe and non-severe crash outcomes at varying
threshold levels, which is paramount in the domain of AV safety. A notable observation from the
precision-recall curve is the high precision that the model maintains even at high recall levels,
7
which is not commonly observed in typical models, especially in imbalanced datasets like AV crash
data.
Figure 4. Precision-recall curve
The shape of the curve also provides valuable information. The initial high precision can
be particularly useful for settings where the cost of false positives is high, but as we move towards
the right on the curve, increasing recall, the precision moderately declines, indicating a trade-off
that must be managed. The specific point on the curve where we decide to operate the model can
be chosen based on the relative costs of false negatives (missed severe crashes) versus false
positives (unnecessary interventions). The curve's progression also presents a nuanced picture of
the classifier's performance across different operational points. In some regions, we observe that
the precision can be recovered slightly after initial drops, which suggests that the classifier's
performance is not monotonically decreasing with respect to recall; rather it exhibits a complex
relationship with the threshold settings. This is indicative of the model's sophistication and its
ability to adapt to the intricacies of the data.
Validation Curve
The validation curve presented in figure 5 further substantiates the efficacy of the model in
predicting AV crash outcomes. The curve plots the training and cross-validation scores as a
function of the number of estimators, providing insight into the model's learning dynamics and its
generalization capabilities. The training score remains high as the number of estimators increases,
suggesting that the model has sufficient capacity to capture the underlying patterns in the training
data. Notably, there is a marginal diminishing return on accuracy as the number of estimators grows
beyond 10, indicating that adding more estimators beyond this point does not significantly enhance
the model's ability to fit the training data. The cross-validation score exhibits a more nuanced
behavior. Starting from a lower bound, the accuracy improves sharply as the number of estimators
increases to 5, signifying that the model benefits from additional complexity up to a certain point.
Beyond this point, however, the curve begins to plateau, with a slight undulation that suggests an
optimal range rather than a singular optimal point. This behavior is characteristic of the bagging
classifier's ability to stabilize its predictions without overfitting the data as more estimators are
introduced, as is evidenced by the convergence of training and cross-validation accuracy.
8
Figure 5. Validation curve
The shaded area representing the variability of the cross-validation score indicates the
confidence interval of the generalization performance. The relatively tight confidence intervals
around the cross-validation score as compared to the training score further emphasize the model's
stability. It is worth noting that the model does not exhibit overfitting, as the training and cross-
validation scores remain close across the spectrum of complexity. This demonstrates its robustness
and ability to generalize to new, unseen data.
Learning Curve
The learning curve presented in figure 6 provides additional depth to our understanding of the
bagging classifier's performance characteristics. It illustrates the relationship between the training
size and the model's score for both the training and cross-validation sets and makes it evident that
the training score starts high (close to 1.00) and remains relatively stable regardless of the increase
in training size. This is indicative of the model's ability to learn effectively from the training data.
The consistently high score across different training sizes underscores the model's low bias and its
ability to capture the underlying data distribution without overfitting, as the training score does not
decrease with the addition of more training data.
The cross-validation score starts at a lower point when the training size is small, suggesting
underfitting where the model is not complex enough to capture the underlying pattern in the data.
As the training size increases, however, the cross-validation score improves significantly,
indicating that with more data, the model is better able to generalize its predictions to unseen data.
The increase in the cross-validation score with training size is an encouraging sign of the model's
learning capability. Notably, there is a dip in the cross-validation score as the training size increases
from 100 to 150. This could be attributed to several factors, such as variability in the data, a
particular subset of data that is more difficult to learn, or simply the randomness inherent in the
cross-validation process. However, as the training size further increases, the cross-validation score
recovers, suggesting that the model is able to overcome these issues with more data.
9
Figure 6. Learning curve
The plateauing of the cross-validation score for larger training sizes suggests that additional
data beyond this point does not result in overfitting. This is a desirable property in a predictive
model, as it indicates that the model has reached a level of complexity sufficient to capture essential
patterns in the data, without being swayed by noise or outliers. The learning curve, therefore,
complements our previous discussion on the precision-recall and validation curves by highlighting
the bagging classifier's consistent performance and its ability to learn effectively as more data is
provided.
CONCLUSION
This study uses a bagging classifier model to present a comprehensive analysis of the prediction
of injuries from crashes involving AVs. The methodology, which encompasses data preprocessing,
model development, and rigorous evaluation metrics, successfully addresses key challenges in AV
crash data analysis, notably class imbalance and data complexity. The choice of a bagging classifier
was a strategic decision that was based on its effectiveness with complex and imbalanced datasets,
and it adeptly managed the data imbalance through class weighting and stratified sampling,
ensuring equitable class representation.
The comprehensive evaluation of the bagging classifier model through various
performance metrics and curves demonstrates its substantial potential for reliably predicting AV
crash outcomes. The balanced accuracy of 0.59, while modest, ensures an equitable consideration
of both severe and non-severe crash predictions, which is crucial for real-world applications where
both classes are of significant interest. The precision of 0.94 and recall of 0.97, as illustrated by
the precision-recall curve, underscore the model's ability to accurately identify severe crashes
while minimizing false positives. This balance is crucial in AV contexts, where the costs of
misprediction can range from unnecessary alarms to missed opportunities for preventing accidents.
The precision-recall curve not only validates the high performance of our bagging classifier, but
also provides strategic guidance on selection of the threshold. The high modified F1 score of 2.39,
which was adjusted for the class imbalance, also reflects the curve's indication that the model
effectively balances the precision and recall, making it a trustworthy tool for critical safety
10
applications. The validation curve suggests that the model is sufficiently complex to capture the
necessary patterns in the data, while avoiding both underfitting and overfitting. The learning curve
further reinforces these findings, showing that the model's ability to generalize improves with more
data but stabilizes, indicating that the model is capturing the essential structure of the problem
without being unduly influenced by noise.
Overall, the ability of the bagging classifier's performance to handle the intricacies of
imbalanced AV crash data with a high degree of precision and recall positions it as a promising
tool for enhancing safety in the burgeoning field of AV s. As autonomous technology continues to
evolve, the deployment of such advanced predictive models will play a pivotal role in mitigating
risks and improving outcomes in AV operations. The insights gained from the various performance
evaluations will guide future refinements of the model, ensuring that they continue to serve the
needs of an industry where the stakes are nothing short of life-altering.
The main limitation of this study is its reliance on geographically limited data from the CA
DMV. Future studies could explore integrating additional data sources, real-time data, and varying
environmental factors to further enhance the model's predictive capabilities. Investigating the
application of the bagging classifier in different geographic and regulatory contexts would provide
a more comprehensive understanding of its effectiveness across diverse environments.
ACKNOWLEDGMENT
We gratefully acknowledge the support and generosity of the Transportation Consortium of South-
Central States (Tran-SET), without which the present study could not have been completed.
DECLARATION OF COMPETING INTEREST
The authors declare that they have no known competing financial interests or personal
relationships that could have appeared to influence the work reported in this paper.
REFERENCES
Boggs, A.M., Wali, B. and Khattak, A.J., 2020. Exploratory analysis of automated vehicle crashes
in California: A text analytics & hierarchical Bayesian heterogeneity-based approach.
Accident Analysis & Prevention, 135, p.105354.
Channamallu, S. S., Kermanshachi, S., & Pamidimukkala, A., 2023. Impact of Autonomous
Vehicles on Traffic Crashes in Comparison with Conventional Vehicles. In International
Conference on Transportation and Development 2023 (pp. 39-50).
Channamallu, S. S., Kermanshachi, S., Rosenberger, J. M., & Pamidimukkala, A., 2023a. A review
of smart parking systems. Transportation Research Procedia, 73, 289-296.
https://doi.org/10.1016/j.trpro.2023.11.920
Channamallu, S. S., Kermanshachi, S., Rosenberger, J. M., & Pamidimukkala, A., 2023b. Parking
occupancy prediction and analysis - a comprehensive study. Transportation Research
Procedia, 73, 297-304. https://doi.org/10.1016/j.trpro.2023.11.921
Das, S., Dutta, A. and Tsapakis, I., 2020. Automated vehicle collisions in California: Applying
Bayesian latent class model. IATSS research, 44(4), pp.300-308.
11
Ding, S., Abdel-Aty, M., Wang, D., Wang, Z., Zheng, O., Barbour, N. and Yuan, R., 2023.
Exploratory Analysis of the Crash Severity between Vehicular Automation (SAE L2-5)
with Multi-Source Data.
Etminani-Ghasrodashti, R., Patel, R., Kermanshachi, S., Rosenberger, J., & Foss, A., 2022a.
Exploring Factors Affecting Willingness to Ride Shared Autonomous Vehicles. In
Proceedings of Transportation Research Board 101 Annual Conference.
Etminani-Ghasrodashti, R., Ketankumar Patel, R., Kermanshachi, S., Rosenberger, J. M., & Foss,
A., 2022b. Modeling users’ adoption of shared autonomous vehicles employing actual
ridership experiences. Transportation Research Record, 2676(11), 462-478.
Etminani-Ghasrodashti, R., Kermanshachi, S., Rosenberger, J. M., & Foss, A., 2023. Exploring
motivating factors and constraints of using and adoption of shared autonomous vehicles
(SAVs). Transportation Research Interdisciplinary Perspectives, 18, 100794.
Favarò, F.M., Nader, N., Eurich, S.O., Tripp, M. and Varadaraju, N., 2017. Examining accident
reports involving autonomous vehicles in California. PLoS one, 12(9), p.e0184952.
Houseal, L.A., Gaweesh, S.M., Dadvar, S. and Ahmed, M.M., 2022. Causes and effects of
autonomous vehicle field test crashes and disengagements using exploratory factor
analysis, binary logistic regression, and decision trees. Transportation research record,
2676(8), pp.571-586.
Khan, M.A., Etminani-Ghasrodashti, R., Kermanshachi, S., Rosenberger, J. M., & Foss, A., 2022a.
Identifying Usage of Shared Autonomous Vehicles (SAVs): Early Findings from a Pilot
Project. In Transportation Research Board 101st Annual Meeting Washington, DC.
Khan, M.A., Etminani-Ghasrodashti, R., Shahmoradi, A., Kermanshachi, S., Rosenberger, J. M.,
& Foss, A., 2022b. Integrating shared autonomous vehicles into existing transportation
services: evidence from a paratransit service in Arlington, Texas. IJCE, 20(6), 601-618.
Khan, M. A., Etminani-Ghasrodashti, R., Kermanshachi, S., Rosenberger, J. M., Pan, Q., & Foss,
A., 2022c. Do ridesharing transportation services alleviate traffic crashes? A time series
analysis. Traffic injury prevention, 23(6), 333-338.
Khan, M.A., Patel, R. K., Pamidimukkala, A., Kermanshachi, S., Rosenberger, J. M., Hladik, G.,
& Foss, A., 2023a. Factors that determine a university community’s satisfaction levels with
public transit services. Frontiers in Built Environment, 9, 1125149.
Khan, M. A., Etminani-Ghasrodashti, R., Kermanshachi, S., Rosenberger, J. M., Pan, Q., & Foss,
A., 2023b. Understanding Students’ Satisfaction with University Transportation. In
International Conference on Transportation and Development 2023 (pp. 522-532).
Kutela, B., Avelar, R.E. and Bansal, P., 2022. Modeling automated vehicle crashes with a focus on
vehicle at-fault, collision type, and injury outcome. Journal of transportation engineering,
Part A: Systems, 148(6), p.04022024.
Leilabadi, S.H. and Schmidt, S., 2019. In-depth analysis of autonomous vehicle collisions in
California. In 2019 IEEE Intelligent Transportation Systems Conference (ITSC) (pp. 889-
893). IEEE.
Novat, N., Kidando, E., Kutela, B. and Kitali, A.E., 2023. A comparative study of collision types
between automated and conventional vehicles using Bayesian probabilistic inferences.
Journal of safety research, 84, pp.251-260.
Pamidimukkala, A., Patel, R. K., Kermanshachi, S., Rosenberger, J. M., & Tanvir, S., 2023a. A
Review on Shared Mobility and Electric Vehicles. In International Conference on
Transportation and Development 2023 (pp. 333-342).
12
Pamidimukkala, A., Kermanshachi, S., & Patel, R., 2023b. An Exploratory Analysis of Crashes
Involving Autonomous Vehicles. In International Conference on Transportation and
Development 2023 (pp. 343-350).
Patel, R. K., Etminani-Ghasrodashti, R., Kermanshachi, S., Rosenberger, J. M., & Foss, A., 2022a.
Exploring willingness to use shared autonomous vehicles. International Journal of
Transportation Science and Technology, 12(3), 765-778.
Patel, R. K., Etminani-Ghasrodashti, R., Kermanshachi, S., Rosenberger, J., & Foss, A. 2022b.
How Riders Use Shared Autonomous Vehicles. In Automated People Movers and
Automated Transit Systems, ASCE International Conference on Transportation &
Development, pp. 81-93.
Patel, R., Channamallu, S. S., Khan, M. A., Kermanshachi, S & Pamidimukkala, A., 2023a. An
Exploratory Analysis of Temporal and Spatial Patterns of Autonomous Vehicle Collisions.
Public Works Management & Policy. DOI: 10.1177/1087724X231217677.
Patel, R. K., Pamidimukkala, A., Kermanshachi, S., & Etminani-Ghasrodashti, R., 2023b. Disaster
preparedness and awareness among university students: a structural equation analysis.
International journal of environmental research and public health, 20(5), 4447.
Ren, W., Yu, B., Chen, Y. and Gao, K., 2022. Divergent effects of factors on crash severity under
autonomous and conventional driving modes using a hierarchical Bayesian approach.
International journal of environmental research and public health, 19(18), p.11358.
Sinha, A., Vu, V., Chand, S., Wijayaratna, K. and Dixit, V., 2021. A crash injury model involving
autonomous vehicle: Investigating of crash and disengagement reports. Sustainability,
13(14), p.7938.
Song, Y., Chitturi, M.V. and Noyce, D.A., 2021. Automated vehicle crash sequences: Patterns and
potential uses in safety testing. Accident Analysis & Prevention, 153, p.106017.
Wang, S. and Li, Z., 2019. Exploring the mechanism of crashes with automated vehicles using
statistical modeling approaches. PloS one, 14(3), p.e0214550.
Xu, C., Ding, Z., Wang, C. and Li, Z., 2019. Statistical analysis of the patterns and characteristics
of connected and autonomous vehicle involved crashes. Journal of safety research, 71,
pp.41-47.
Ye, W., Wang, C., Chen, F., Yan, S. and Li, L., 2021. Approaching autonomous driving with
cautious optimism: analysis of road traffic injuries involving autonomous vehicles based
on field test data. Injury prevention, 27(1), pp.42-47.
Yuan, Q., Xu, X., Wang, T. and Chen, Y., 2022. Investigating safety and liability of autonomous
vehicles: Bayesian random parameter ordered probit model analysis. Journal of intelligent
and connected vehicles, 5(3), pp.199-205.
Zhu, S. and Meng, Q., 2022. What can we learn from autonomous vehicle collision data on crash
severity? A cost-sensitive CART approach. Accident Analysis & Prevention, 174,
p.106769.
