Forty-four students presented original work at the Portland State University (PSU) Summer Research Symposium on August 15, and six of those students chose to focus on transportation topics.
Naomi Cai, River Johnson, Danielle Justo Olivia Wang, Paris Wu, and Rayna Yu all devoted their work during a ten-week research program to solving transportation challenges with the use of advanced computational techniques. Each student was advised by a PSU faculty member.
The National Science Foundation (NSF)-funded Research for Undergraduates (REU) and altREU programs are managed by PSU's Teuscher Lab, a lab focused on next generation computing models and architectures led by professor Christof Teuscher of the Maseeh College of Engineering and Computer Science. The programs are designed for curious, motivated students from any university who are interested in designing, programming, and using computers to benefit society. Students select topics they are interested in and work with their faculty advisors to develop a unique research program and put their knowledge into practice.
"I'm always impressed by what undergraduate students can learn and achieve in just one summer. Perhaps the most valuable lesson they discover is that research is not a straight line: it requires tremendous patience and resilience. In research, many attempts don't work out, leading to frequent setbacks. Yet it's precisely these setbacks that ultimately pave the way to breakthrough discoveries," Teuscher said.
To see the variety of topics investigated by students this summer, check out all the final presentations. Learn more about the transportation projects below, and join us in congratulating these students on their work improving the future of transportation!
Naomi Cai, Washington University in St. Louis
Advisor: Sirisha Kothuri
Topic: Improving Pedestrian Count Estimation with Machine Learning and Data Fusion
Naomi's research explored improving pedestrian count estimation by applying machine learning techniques to Strava fitness app data, along with static and count data. Her initial experiments revealed that model performance was heavily dependent on data splits, leading her to adopt leave-one-out cross-validation for more stable results. Ultimately, Naomi found that CatBoost and Random Forest models generalized better to higher-count data, indicating a trade-off between accuracy at low versus high counts, and identified speed limit as a top-ranked feature in her analysis. By improving estimation models with machine learning and novel data sources, Naomi’s work can help develop cost-effective ways to understand where people are walking.
"I hope I can do more work on the topic, since I thought it was pretty interesting. I think I can definitely improve on the work I did if I'm able to find more data, perhaps from different cities, to work with," Naomi said.
Watch Naomi's Summer Research Symposium presentation to learn more.
River Johnson, Western Carolina University
Advisor: Tammy Lee
Topic: Comparing Ultralytics YOLOv8 and YOLOv10 for Multimodal Transportation Counts
River's presentation focused on improving multimodal transportation counts using Ultralytics YOLO V8 and YOLO V10 models to better inform pedestrian-friendly urban design. They used a Python script written by Alicia Hopper, an altREU program participant from 2024 who also worked with Tammy Lee. The project aimed to overcome the limitations of traditional manual or specialized device counting methods by leveraging object detection models on video footage. Using the Euro City Persons data set, River found that YOLO V8 generally performed slightly better, particularly for pedestrians and bicyclists, while YOLO V10 was superior for categories with fewer samples, such as wheelchair users.
"It was challenging, in a really constructive way. I learned a lot in just 10 weeks, and it was mostly self-directed, but whenever I got stuck on something there was still plenty of support available," River said.
River also presented this work to the PORTAL Users Group (PUG) on August 7. PORTAL is the official transportation data archive for the Portland-Vancouver Metropolitan region.
"It's really cool that students in this program can build on one another's work from year to year and make it their own. Allie Hopper developed the Python script in 2024, and trained a computer model to predict nonmotorized counts using images from a EuroCity Persons Dataset. This year, River was able to focus on improving the accuracy and versatility of this method. It's wonderful to see how each student tackles a problem differently and what other ideas they can come up with," said advisor Tammy Lee, TREC's Transportation Data Program Manager.
Watch River's Summer Research Symposium presentation to learn more.
Danielle Justo, Smith College
Advisor: Banafsheh Rekabdar
Topic: Exploring Detection Methods for Adversarial Attacks on Multimodal RL Agents
Danielle's presentation delved into exploring adversarial attack detection methods for multimodal reinforcement learning (RL) agents. While her work is not necessarily specific to transportation, it can be used in a variety of contexts. Advisor Banafsheh Rekabdar conducts research in bicycle and pedestrian detection, among other areas.
Adversarial attacks involve subtle changes to input data that cause models to behave unexpectedly, often leading to misclassification. Danielle's objectives included evaluating detection as an adversarial defense and examining how different combinations of attacked modalities influenced detection effectiveness. She trained a baseline RL agent in the MuJoCo Ant Maze environment and attacked its observation space using the Fast Gradient Sign Method (FGSM). Her experiments with various detection models—including clustering, shallow classifiers, and neural networks—showed that classifiers performed better overall than clustering methods. Notably, angular attacks were more detectable than velocity attacks, and a fully connected neural network (FCN) emerged as the best-performing detector, demonstrating a strong importance towards angular modalities. This research highlighted the disproportionate effect of attacks on different modalities and suggested that adversaries must balance the impact on model performance with detectability.
Watch Danielle's Summer Research Symposium presentation to learn more.
Olivia Wang, Tufts University; Paris Wu, Cornell University; Rayna Yu, Northeastern University
Advisor: Christof Teuscher
Topic: NavigAid
Olivia, Paris, and Rayna presented NavgiAid, a route analysis and navigation model designed to improve pedestrian safety in Boston, addressing the limitations of existing car-centric navigation systems. Their objective was to identify the safest walking routes by integrating features linked to crash risk, including speed limits, crosswalk coverage, lighting, sidewalk width, slope, tree coverage, and pedestrian ramps.
Using the Open Route Service API, their system generated three candidate routes between user-defined points and flagged safety features along each using Analyze Boston datasets. These features were then processed with a Random Forest classifier trained on 300,000 labeled points with Vision Zero crash data, achieving 94% accuracy and an AUC of 0.99. Speed limit emerged as the most influential factor, with indirect features like lighting and trees contributing less. The resulting app displayed routes ranked by safety scores, highlighting flagged features with color-coded visualizations and allowing users to filter preferences. Future directions included refining route comparisons, expanding to other cities with quality data, and leveraging outputs to identify inequities in pedestrian infrastructure for policy and investment decisions.
"In the course of working together in the altREU program this summer to build NavigAid, a navigation model that is focused on pedestrian safety, we learned a lot about both the research process and the issue of pedestrian accessibility, accident risk factors, and infrastructure in our local neighborhood of Boston. We greatly appreciated the opportunity to present and get feedback on our work from experienced researchers in the field, and to learn from other students in our cohort. Their insights helped us to continuously improve our parameters for pedestrian safety as well as the clarity of our research communication," said Olivia, Rayna, and Paris.
Watch their presentation to learn more.
Header: Image from IEEE Transactions on Pattern Analysis and Machine Detection
Image from Morgunov, Anton. "Object Detection with YOLO: Hands-On Tutorial" April 2025
Image courtesy of Sirisha Kothuri
Portland State University's Transportation Research and Education Center (TREC) is a multidisciplinary hub for all things transportation. We are home to the Initiative for Bicycle and Pedestrian Innovation (IBPI), the data programs PORTAL and BikePed Portal, the Better Block PSU program, and PSU's membership in PacTrans, the Pacific Northwest Transportation Consortium. Our continuing goal is to produce impactful research and tools for transportation decision makers, expand the diversity and capacity of the workforce, and engage students and professionals through education, seminars, and participation in research. To get updates about what's happening at TREC, sign up for our monthly newsletter or follow us on social media.
