Trajectory Prediction for Autonomous Vehicles in Construction Zones*
Session Number
3
Advisor(s)
Dr. John M. Dolan, Carnegie Mellon University; Juan Alvarez-Padilla, Massachusetts Institute of Technology
Location
A115
Discipline
Computer Science
Start Date
15-4-2026 2:15 PM
End Date
15-4-2026 3:00 PM
Abstract
Autonomous driving technologies have the potential to revolutionize transportation by reducing accidents, improving traffic efficiency, decreasing fuel consumption, and increasing mobility. An autonomous vehicle’s ability to generate a safe future trajectory depends on several subsystems, including environment perception, intention prediction, trajectory prediction, and planning. However, many of these subsystems are challenged in construction zones, where environments are complex, cluttered, and highly uncertain. Traffic laws may not fully apply, road markings may be missing or misleading, lanes may vary in width, and unexpected objects may appear. This study addresses the trajectory prediction subsystem by augmenting the prediction pipeline with a discrete encoder that performs attention over objects related to construction zones, using construction data where available to improve performance. The enhanced model uses object annotations to encode agents of the “movable-object” class that are not already used in the model.
Trajectory Prediction for Autonomous Vehicles in Construction Zones*
A115
Autonomous driving technologies have the potential to revolutionize transportation by reducing accidents, improving traffic efficiency, decreasing fuel consumption, and increasing mobility. An autonomous vehicle’s ability to generate a safe future trajectory depends on several subsystems, including environment perception, intention prediction, trajectory prediction, and planning. However, many of these subsystems are challenged in construction zones, where environments are complex, cluttered, and highly uncertain. Traffic laws may not fully apply, road markings may be missing or misleading, lanes may vary in width, and unexpected objects may appear. This study addresses the trajectory prediction subsystem by augmenting the prediction pipeline with a discrete encoder that performs attention over objects related to construction zones, using construction data where available to improve performance. The enhanced model uses object annotations to encode agents of the “movable-object” class that are not already used in the model.