Interaction-Aware Personalized Vehicle Trajectory Prediction Using Temporal Graph Neural Networks

Abdelraouf, Amr; Gupta, Rohit; Han, Kyungtae · 2023 · Crossref

DOI: 10.1109/itsc57777.2023.10421928

archive: archived pipeline: cataloged verified

Get this paper ↗ (DOI — opens at the source; we link to it, we don't host it)

Summary

This paper addresses the lack of personalized, interaction-aware vehicle trajectory prediction methods, a gap that limits the effectiveness of Advanced Driver Assistance Systems (ADAS) and autonomous vehicles. While existing approaches rely on generic models trained on large datasets, they often overlook individual driving styles and the complex spatio-temporal interactions between a target vehicle and its surroundings. The authors propose a novel framework that combines Graph Convolution Networks (GCN) and Long Short-Term Memory (LSTM) networks to capture these interactions, utilizing transfer learning to personalize predictions for specific drivers. To overcome the scarcity of long-term, naturalistic driving data required for such modeling, the researchers conducted a human-in-the-loop simulation using the CARLA driving simulator. Five drivers provided 40 minutes of naturalistic driving data each under varying traffic densities (100, 200, and 300 background vehicles) on a simulated highway. The methodology involves a two-step transfer learning process: first, a base GCN-LSTM model is pre-trained on the large-scale CitySim dataset to learn generic interaction patterns; second, the model is fine-tuned for each driver using their specific trajectory data. During fine-tuning, the encoder weights remain frozen to preserve learned spatial features, while the decoder is updated to adapt to individual driving characteristics. Experimental results demonstrate that the personalized GCN-LSTM model significantly outperforms generic counterparts, particularly for longer prediction horizons. Compared to a generic GCN-LSTM, the personalized model achieved Root Mean Square Error (RMSE) reductions ranging from 2% at 1-second horizons to 43% at 5-second horizons. The personalized approach also surpassed individual models trained exclusively on driver-specific data without pre-training, indicating that pre-training prevents overfitting and enhances generalization. Furthermore, increasing the duration of driving data used for fine-tuning from 5 to 30 minutes consistently reduced prediction errors, with the most significant improvements observed at longer horizons. The study concludes that incorporating personalization into interaction-aware trajectory prediction enhances accuracy and reliability, which is critical for safety-critical applications like collision warning and automated braking. By reducing false positives and improving prediction precision, this approach can increase driver trust in ADAS and facilitate cooperative behaviors in mixed traffic through Vehicle-to-Vehicle communication. The findings underscore the importance of leveraging large-scale pre-training combined with individual fine-tuning to effectively capture unique driver behaviors and complex traffic interactions.

Provenance

The full processing record for this entry. Every stage of this paper's journey through the pipeline is logged — what ran, with which tool and model, how many attempts it took, and when it last completed.

Summary generated by qwen3.6-27b-prismaquant on 2026-06-26; verification: verified.

Topics

Ranked by relevance to this paper. Hover a topic for its definition.

• anticipation