Meet the TRANSFOR 22 Finalists Analyzing 3D LiDAR Data to Make Intersections Safer
Updated: Apr 28
The Transportation Forecasting Competition (TRANSFOR 22) sought creative and innovative approaches to improving vulnerable road user safety at intersections. Unfortunately, one pedestrian is killed by a vehicle every 85 minutes in the United States, according to the US Department of Transportation, with the disabled and elderly road users at heightened risk. Better insights into vehicle and pedestrian interactions, coupled with a more dynamic and responsive traffic system, can help prevent these accidents – creating safer communities for all. But at the foundation, any safety initiative requires accurate and robust data.
The three finalists from the University of Michigan, University of Washington, and Korea Advanced Institute of Science and Technology (KAIST) were provided with the same LiDAR data collected from the MLK Smart Corridor, an urban testbed developed by the Center for Urban Informatics and Progress (CUIP) at the University of Tennessee at Chattanooga (UTC). The dataset was collected by Ouster digital LiDAR sensors and processed with Seoul Robotics’ 3D perception software, labeling objects as cars, pedestrians, cyclists, or miscellaneous.
Participants were then tasked with designing a unique solution that would improve road user safety by evaluating the accuracy of sub-classifying higher-risk pedestrians and predicting the time needed for a pedestrian to cross the street.
One Challenge, Three Solutions
The first place winner from the University of Michigan Transportation Research Institute began their project by processing the data to reduce noise and improve accuracy. To do this, the team mapped the data into a grid (resembling a crosswalk) that was then divided into four quadrants. Based on the trajectories, they could filter outlier vehicles and pedestrians from the dataset. Once the team incorporated signal information and determined the medoid of clustered car trajectories, they could begin addressing the challenge at hand.
To quantify the pedestrian risk of crossing the street, the team developed a time-to-collision (TTC) calculation that determined the probability of risk depending on the direction a vehicle was traveling based on its clustered trajectory. For the pedestrian trajectories, the team calculated TTC using linear least squares and three different estimators, choosing the one determining the highest risk. The team then used motion-based and machine learning models to determine the amount of time a pedestrian would require to cross the street. Ultimately, the Long Short-Term Memory method was most accurately able to determine the amount of time required to cross and whether the signal allotted enough time to do so safely.
The University of Washington Star Lab began their project with an exploratory analysis to ensure that the IDs provided in the initial dataset were accurate. From there, the team filtered the data based on location, velocity, and movement to remove any outlying non-pedestrian data. With this refined dataset in place, the team used time series clustering to identify high-risk pedestrians based on their velocity over time. They then constructed an early detection algorithm that could alert a crossing signal early on if a pedestrian requires more time based on walking speed.
KAIST, the third place winner, used a combination of machine learning and deep learning to solve the challenge. To sub-classify pedestrians, the team used Support Vector machines, leveraging machine learning algorithms to filter the data and determine if a pedestrian was crossing the street on foot or in a wheelchair, based on velocity and height with 93% accuracy. To create an arrival-time prediction model, the team tested Deep Neural Networks (DNN), Long Short Term Memory (LSTM), Gated Recurrent Unit (GRU), and Transformer to accurately predict pedestrians’ arrival time with 1.2 seconds error margin.
Making TRANSFOR 22 Possible
All three finalists presented their projects during the Transportation Research Board (TRB) Annual Meeting last month in Washington, D.C. Final rankings were determined based on classification accuracy, novelty of solutions, quality of code, quality of paper and oral presentations.
TRB is a division of the National Academy of Sciences, Engineering, and Medicine, dedicated to independent research and advice on all modes of transportation. Founded in 1920, TRB has provided trusted, impartial, evidence-based research and information exchange to inform public policy decisions.
AED50 Artificial Intelligence and Advanced Computing Applications, the organizers of this program, would like to thank its partners and sponsors for their participation. The IEEE ITSS Technical Activities Sub-Committee “Smart Mobility and Transportation 5.0” was a key supporter, while the Center for Urban Informatics and Progress (CUIP) at The University of Tennessee at Chattanooga (UTC), NSF, City of Chattanooga, Seoul Robotics and Ouster LiDAR were proud sponsors and contributors.
The organizers and contributors extend their gratitude to all three finalist teams for presenting their projects during the TRB meeting, and to all the groups who submitted their projects to be considered for this Hackathon.