Back to search

University of Minnesota

Lane-changing prediction using GNN-LSTM architecture based on surrounding vehicle behavior

Abstract

dc:description.abstract

Lane changes, particularly on highways, contribute significantly to road accidents, traffic breakdowns, and speed fluctuation, making them a critical factor in traffic safety and efficiency. Understanding and predicting these maneuvers is essential for developing advanced driver assistance systems and autonomous vehicles. This underscores the fundamental importance of understanding and forecasting the intentions behind these maneuvers with sufficient lead time to enable proactive safety interventions. This research addresses the lane-change prediction problem as a multi-class classification task, with labels indicating the probability of a driver making a lane change (left or right) or maintaining lane position based on influencing features in the upcoming seconds. We present an end-to-end Graph Neural Networks-Long Short-Term Memory (GNN-LSTM) hybrid deep learning model for predicting lane-change maneuvers from trajectory data using comprehensive features of surrounding vehicles, with prediction horizon times extending up to 5 seconds ahead. The proposed GNN-LSTM architecture combines the spatial relationship modeling capabilities of graph neural networks with the temporal sequence processing strengths of LSTM networks. The GNN-LSTM model is built and validated using a novel dataset from the Interstate-24 Mobility Technology Interstate Observation Network (I-24MOTION), which provides high-resolution trajectory data from real-world highway scenarios. Comprehensive experimental evaluation demonstrates that the proposed model can efficiently predict lane changes along with their specific directions (left, right, or lane keeping). The model achieves a high overall accuracy of 93.6%. The analysis also reveals the significant impact of surrounding vehicle features on lane-changing maneuvers through attention mechanism visualization, showing that the model learns to focus on physically meaningful vehicle interactions such as gap assessment. This research makes several key contributions to the field: (1) development of a hybrid GNN-LSTM architecture specifically designed for lane-changing intention prediction, (2) comprehensive feature engineering that captures inter-vehicle relationships, (3) extensive validation on high-quality real-world highway data, and (4) systematic sensitivity analysis of input time windows and prediction horizon parameters to optimize model performance and establish optimal temporal configurations for lane change prediction. The findings pave the way for enhanced highway safety systems, improved driving behavior modeling, and significant advancements in autonomous vehicle technology and traffic management systems.

Author and committee

dc:creator, dc:contributor.*
Author dc:creator
  • Nouhi, Shaghayegh

Subjects

dc:subject × 4

Rights

Language dc:language.iso
en

Identifiers

dc:identifier.*
Dc Identifier Citation
Nouhi, Shaghayegh. (2025). Lane-changing prediction using GNN-LSTM architecture based on surrounding vehicle behavior. Retrieved from the University Digital Conservancy, https://hdl.handle.net/11299/280267.

Chain of custody

source
Harvested from
University of Minnesota
Base URL
conservancy.umn.edu/server/oai/request
Last updated
2026-11-19
Source record
OAI-PMH GetRecord
citation

Nouhi, Shaghayegh. Lane-changing prediction using GNN-LSTM architecture based on surrounding vehicle behavior. 2025. https://hdl.handle.net/11299/280267