Sunday, Aug 3: 4:00 PM - 5:50 PM
0263
Invited Paper Session
Music City Center
Room: CC-202A
With rapid advances in modern technology, artificial intelligence has become a fast-
growing area in the transportation research field. The massive data collected from
advanced sensing technologies can assist model training for artificial intelligence and
hence improve safety, equity and sustainability for transportation systems. The
proposed session will introduce to the audience some recent advances in artificial
intelligence methodologies in transportation research. The method development is
highly motivated by the real problems raised by the automotive industry and
transportation-related government agencies.
The invited presentations are related to a few broader topics in artificial intelligence in
transportation research. Dr. Naveen Eluru focuses on analyzing the decision process
using real-world data. He develops econometric models to provide recommendations to
policymakers for improving the transportation system. Dr. Wenlong Jin explores
potential application of large language models in transportation systems trained by
large-scale transportation datasets. Dr. Gaurav Pandey will present an emerging AI
method to create accurate and visually plausible representations of the environment
using LIDAR and camera data for 3D mapping and visual localization. Dr. Arpan Kusari
develops a simulated physical process for generating broken lens as a class of physics-
based adversarial samples. The objective is to provide a robust physics-based process
for generating adversarial samples for safety in object detection.
By appropriately leveraging the transportation data, these novel methods coupled with
sound artificial intelligence methodologies provide great potentials for improvements in
areas such as transportation policy making, sensing system design, vehicle relocation,
and autonomous driving safety. This session will make a great opportunity to learn more
about such advances in artificial intelligence in transportation research and initiate
discussions for practitioners working in the transportation areas and/or with applied
interests in both academia and industry.
Applied
Yes
Main Sponsor
Transportation Statistics Interest Group
Presentations
Independent traffic crash modeling approaches do not account for the embedded relationships
related to the multi-resolution data structure, leading to mis-specified estimations. The recently
developed integrated frameworks demonstrate the capability of addressing this drawback. The
current study proposes an integrated framework that accommodates information from multiple
spatial units and observation resolutions. Specifically, the study develops an integrated model
system that allows for the influence of independent variables from disaggregate crash record,
micro-facility (segment and intersection) and macro (traffic analysis zone) level simultaneously
within the macro level propensity estimation. The empirical analysis considers disaggregate
crash records of 1,818 segments and 4,184 intersections from 300 traffic analysis zones in the
City of Orlando, Florida. These crash records contain crash-specific factors, driver and vehicle
factors, roadway, road environmental and weather information of each crash record. For micro-
facility and macro levels, an exhaustive set of independent variables including roadway and
traffic factors, land-use and built environment attributes, and sociodemographic characteristics
are considered. The proposed model system can also accommodate for hierarchical correlations
among the data across observation resolutions and parameter variability across the system. The
empirical analysis is augmented by employing several goodness of fit and predictive measures.
The results clearly demonstrate the improved performance offered by the proposed integrated
model system relative to the non-integrated model. A validation exercise also highlights the
superiority of the proposed framework. The application of the proposed integrated framework
can allow transportation professionals to adopt policy-based, site-specific, and outcome-specific
solutions simultaneously.
While established econometric approaches use latent variables to model attitudes and improve model fit, we propose a prompt-based framework that uses large language models (LLMs) to provide additional insights into complex reasoning processes surrounding smart charging adoption. Our approach analyzes structured survey profiles to infer behavioral reasoning about smart charging interest and topically relevant attitudes (e.g., privacy, cost, and trust). We evaluate three prompting strategies— zero-shot, chain-of-though, and self-consistency—to assess LLM output fairness across race, income, age and so on, using demographic parity, total variation distance, equalized odds, and equality of opportunity. Early findings indicate that while LLMs produce more neutral outputs than human survey responses at the population level, certain prompting strategies can amplify subgroup disparities . These results caution against assuming moderation implies fairness and highlights the need for multi-level equity diagnostics in human-centered predictive modeling for energy policy. Ongoing work explores causal prompt design and uncertainty-aware inference to improve interpretability and policy sensitivity.
In this talk, Dr. Pandey will present a novel system designed for 3D mapping and visual localization using 3D Gaussian Splatting. The proposed method uses LiDAR and camera data to create accurate and visually plausible representations of the environment. By initiating the training of the 3D Gaussian Splatting map with LiDAR data, the system is able to construct detailed and geometrically precise maps, addressing common issues such as excessive memory usage and imprecise geometry. This preparation makes the method well-suited for visual localization tasks, enabling efficient identification of correspondences between the query image and the rendered image from the Gaussian Splatting.
Digital twins are becoming increasingly important due to their ability to enhance operational efficiency while reducing costs. In the first part of my talk, I will focus on how we can generate digital twins of fractures in the camera or the camera enclosure itself and then overlay them on existing open-source datasets using realistic visualization. We compare our generated results to real broken glass images and show that they are distributionally similar. We also provide the adversarial effect of such generated images, when overlaid on existing datasets, paving a way to provide natural perturbations on images and learning from them.
In the second part of my talk, I will pivot towards using digital twin data in supervised object detection algorithms. The labeling operation required for creating ground truth data makes it a very labor intensive process. Alternatively, we focus on sim-to-real transfer - learning from simulation, and transferring onto a real environment. We perform robust object detection on a collected dataset, without having to label any instances in the target dataset. We show the results from the process, including the robustness of such object detection approaches as compared to learning from labeled examples.