GISNet: Graph-Based Information Sharing Network For Vehicle Trajectory Prediction

Ziyi Zhao, Haowen Fang, Zhao Jin, Qinru Qiu

Abstract

The trajectory prediction is a critical and challenging problem in the design of an autonomous driving system. Many AI-oriented companies, such as Google Waymo, Uber and DiDi, are investigating more accurate vehicle trajectory prediction algorithms. However, the prediction performance is governed by lots of entangled factors, such as the stochastic behaviors of surrounding vehicles, historical information of self-trajectory, and relative positions of neighbors, etc. In this paper, we propose a novel graph-based information sharing network (GISNet) that allows the information sharing between the target vehicle and its surrounding vehicles. Meanwhile, the model encodes the historical trajectory information of all the vehicles in the scene. Experiments are carried out on the public NGSIM US-101 and I-80 Dataset and the prediction performance is measured by the Root Mean Square Error (RMSE). The quantitative and qualitative experimental results show that our model significantly improves the trajectory prediction accuracy, by up to 50.00%, compared to existing models.

Introduction

Predicting the future path of a vehicle is a complex task because drivers do not operate in isolation. Their movements are influenced by a multitude of external factors, including the maneuvers of preceding vehicles, the proximity of tailgaters, and the presence of slow-moving traffic in adjacent lanes. Traditional autonomous driving systems (ADS) often struggle with this "wicked" problem due to the entangled nature of historical paths, relative spatial positions, and the inherently stochastic behavior of surrounding drivers.

Conceptual overview of vehicle trajectory prediction, showing historical paths and multiple predicted possibilities.

Existing approaches typically rely on vanilla LSTMs, which ignore social context, or grid-based social pooling. The latter often uses rigid 2D grids that suffer from computational inefficiency in empty spaces and lose resolution through pooling layers. To address these limitations, we introduce GISNet (Graph-Based Information Sharing Network), which models traffic as a dynamic social network rather than a collection of independent objects.

Methodology

GISNet is an end-to-end differentiable model designed to capture the "latent intentions" of drivers through a formal information-sharing framework. The architecture consists of four primary stages:

  1. Vehicle Information Embedding: Long Short-Term Memory (LSTM) networks encode the 3-second historical trajectory of every vehicle in the scene into a feature vector ($x_{traj}$).
  2. Social Tensor Construction: These vectors are organized into a 13x3 "Social Tensor" to preserve spatial relationships. Convolutional layers (SConv) then extract spatial social features ($x_{social}$).
  3. Graph-Based Information Sharing (GIS) Module: This core component treats vehicles as nodes in a graph. A two-layer Graph Convolutional Network (GCN) allows features to propagate between nodes, capturing non-Euclidean relationships and the "latent intention" of neighbors ($x_{info}$).
  4. Trajectory Generation: The model fuses the vehicle's own history, spatial social context, and GCN-derived interaction data. An LSTM-based decoder then predicts the $x,y$ coordinates for the subsequent 5 seconds.
The GISNet architecture, illustrating the flow from trajectory embedding to graph-based information sharing and final prediction.

To enhance training efficiency, we utilize a batch-wise GCN operation that processes multiple traffic scenes simultaneously using a block-diagonal adjacency matrix.

Experimental Setup and Results

The model was evaluated using the NGSIM (Next Generation Simulation) dataset, specifically the US-101 and I-80 highway segments. The data was partitioned into 8-second windows (3s history, 5s prediction).

Quantitative Performance

GISNet demonstrated significant improvements over baselines such as Constant Velocity (CV), GAIL-GRU, and Multi-Agent Tensor Fusion (MATS). Notably, GISNet achieved a 50% reduction in prediction error at the 1-second horizon and a 21.79% improvement at the 5-second horizon compared to MATS. The Root Mean Square Error (RMSE) for a 5-second prediction was reduced to 3.23 meters, a substantial improvement over the 6.27 meters observed with vanilla LSTMs.

Qualitative Insights

Beyond metrics, GISNet excels in complex scenarios. Visualization shows that the model accurately predicts trajectories in diverse conditions. In high-speed traffic, predictions closely align with ground truth. In congested environments, the model correctly identifies when a vehicle is constrained by its neighbors, accurately predicting stationary behavior or complex lane-change maneuvers that grid-based models often miss.

Conclusion

GISNet represents a shift from viewing traffic as a grid to viewing it as a relational graph. By allowing vehicles to "share" information across a non-Euclidean structure, the model effectively emulates how human drivers surmise the intentions of others. Future work will focus on dynamic graph topologies to better reflect the fluid nature of highway traffic over time.

Citation: Zhao, Z., Fang, H., Jin, Z., & Qiu, Q. (2020). GISNet: Graph-Based Information Sharing Network For Vehicle Trajectory Prediction. arXiv preprint arXiv:2003.11973.

Summary updated

Abstract

The trajectory prediction is a critical and challenging problem in the design of an autonomous driving system. Many AI-oriented companies, such as Google Waymo, Uber and DiDi, are investigating more accurate vehicle trajectory prediction algorithms. However, the prediction performance is governed by lots of entangled factors, such as the stochastic behaviors of surrounding vehicles, historical information of self-trajectory, and relative positions of neighbors, etc. In this paper, we propose a novel graph-based information sharing network (GISNet) that allows the information sharing between the target vehicle and its surrounding vehicles. Meanwhile, the model encodes the historical trajectory information of all the vehicles in the scene. Experiments are carried out on the public NGSIM US-101 and I-80 Dataset and the prediction performance is measured by the Root Mean Square Error (RMSE). The quantitative and qualitative experimental results show that our model significantly improves the trajectory prediction accuracy, by up to 50.00%, compared to existing models.
BibTeX
@inproceedings{zhao2020gisnet,
  author = {Ziyi Zhao and Haowen Fang and Zhao Jin and Qinru Qiu},
  title = {{GISNet: Graph-Based Information Sharing Network For Vehicle Trajectory Prediction}},
  booktitle = {IJCNN},
  year = {2020},
  url = {https://openreview.net/forum?id=FBGw6KBUX9},
}
Figure