Trajectory Prediction for Autonomous Driving Using Agent-Interaction Graph Embedding

Jilan Samiuddin, Benoit Boulet, Di Wu

Introduction

Predicting the future trajectories of surrounding vehicles is a fundamental requirement for safe and efficient autonomous driving systems (ADS). While traditional kinematic models, such as Kalman filters, provide computational efficiency, they often fail in long-term prediction scenarios because they treat vehicles as isolated entities, ignoring the complex "social" interactions inherent in traffic. Conversely, many existing deep learning approaches struggle to effectively bridge the gap between spatial interactions (how vehicles relate to neighbors at a single moment) and temporal dependencies (how those relationships evolve over time).

This paper introduces AiGem (Agent-Interaction Graph Embedding), a lightweight deep learning framework designed to address these challenges. By modeling traffic as a heterogeneous spatial-temporal graph and utilizing Graph Attention Networks (GATs), AiGem achieves state-of-the-art accuracy on long-term (5-second) horizons while maintaining a significantly smaller memory footprint than competing models.

The AiGem Framework

AiGem approaches trajectory prediction through a specialized four-stage pipeline designed to capture the nuances of vehicle movement and interaction.

1. Heterogeneous Graph Formulation

The core innovation of AiGem lies in its graph structure. Rather than a simple homogeneous graph, the model constructs a heterogeneous graph where nodes represent vehicles. Two distinct types of edges are defined:

  • Spatial Edges: These connect vehicles within a specific radius (e.g., 50 meters) at a single time step, capturing immediate social pressure and proximity.
  • Temporal Edges: These connect the same vehicle across consecutive time steps, capturing its momentum, historical intent, and velocity profile.

2. Graph Encoder (Depthwise GAT)

The graph is processed through a series of Graph Attention Network (GAT) layers. Unlike standard convolutions, GATs allow the model to dynamically "attend" to the most relevant neighbors. For instance, the model can learn to prioritize a vehicle merging into the ego car's lane over a vehicle traveling at a constant speed in a distant lane.

3. Sequential GRU Decoder

The encoded graph embedding for the current timestamp is passed to a Gated Recurrent Unit (GRU). The GRU functions as a decoder, generating a sequence of future states based on the rich spatial-temporal features extracted by the encoder.

4. MLP Output Layer

Finally, a Multilayer Perceptron (MLP) predicts the relative change in position ($\Delta Y$) for each time step. Predicting relative displacement rather than absolute coordinates enhances the model's robustness across diverse road geometries and coordinate frames.

Experimental Setup and Results

Datasets and Methodology

The model was evaluated using the industry-standard NGSIM (Next Generation Simulation) datasets, specifically the US-101 and I-80 highway segments. The task involved using 3 seconds of historical data to predict the trajectory for the subsequent 5 seconds. To ensure a fair comparison with existing baselines, raw data was downsampled to 0.2-second intervals.

Performance and Efficiency

AiGem demonstrates significant improvements in long-term forecasting, a domain where most models typically experience a sharp performance degradation. At the 5-second prediction mark, AiGem achieved a Root Mean Square Error (RMSE) of 1.95m, representing a 33.7% improvement over the second-best baseline (GSTCN, which recorded 2.98m). Across all tests, AiGem maintained the lowest Average Displacement Error (ADE) of 1.42m and Final Displacement Error (FDE) of 2.57m.

Beyond accuracy, AiGem is remarkably efficient. With only 74.5K parameters, it is approximately 85% smaller than the popular GRIP++ model (496K parameters). This efficiency makes it an ideal candidate for deployment on edge hardware within autonomous vehicles where computational resources are constrained.

Key Insights

  • Spatial Sensitivity: Ablation studies revealed a "social sweet spot" for interaction. A spatial interaction threshold ($d_{min}$) of 25 meters yielded the best results; beyond this distance, additional vehicle data introduced noise that degraded prediction accuracy.
  • Leader-Follower Dynamics: The model showed higher accuracy in predicting the trajectories of vehicles behind the ego car compared to those in front. This suggests the model successfully learned that a vehicle's future path is heavily dictated by the "leaders" in its immediate forward path.
  • Integrated Modeling: By embedding both spatial and temporal edges directly into the graph, the encoder learns complex "moving interactions" holistically, removing the need for separate, computationally expensive modules for space and time.

Limitations and Future Work

Despite its performance, AiGem has areas for further refinement. Currently, the model predicts trajectories for each actor sequentially, which may present a bottleneck in high-density environments. Additionally, as the prediction horizon extends, the model shows signs of overfitting to the specific driving patterns of the NGSIM dataset. Future iterations could benefit from incorporating explicit road geometry—such as lane boundaries and traffic signs—to further refine predictions in complex urban environments where highway-based kinematic patterns may not apply.

Summary updated

BibTeX
@inproceedings{samiuddin2024trajectory,
  author = {Jilan Samiuddin and Benoit Boulet and Di Wu},
  title = {{Trajectory Prediction for Autonomous Driving Using Agent-Interaction Graph Embedding}},
  booktitle = {ITSC},
  year = {2024},
  doi = {10.1109/itsc58415.2024.10920250},
}
Figure