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

Ziyi Zhao, Haowen Fang, Zhao Jin, Qinru Qiu

Predicting where a car will be five seconds from now sounds simple until you realize that drivers don't move in a vacuum. They react to the car merging ahead of them, the tailgater behind them, and the slow-moving truck in the adjacent lane. In a new paper, researchers from Syracuse University introduce GISNet (Graph-Based Information Sharing Network), a model that treats traffic not as a collection of independent objects, but as a dynamic social network.

TL;DR

GISNet improves vehicle trajectory prediction by using Graph Convolutional Networks (GCNs) to model the "latent intentions" and interactions between a target vehicle and its neighbors. By allowing vehicles to "share" information across a non-Euclidean graph structure, the model reduces prediction errors by up to 50% compared to existing state-of-the-art methods.

What Problem Does This Paper Address?

Most autonomous driving systems (ADS) struggle with trajectory prediction because it is a "wicked" problem involving entangled factors: a driver’s historical path, their relative position to others, and the stochastic (random) behaviors of surrounding vehicles.

Existing solutions often fall into two traps:

  1. Vanilla LSTMs: These look only at a single vehicle's history, ignoring the "social" context.
  2. Grid-based Social Pooling: These models (like CS-LSTM) place vehicles into a rigid 2D grid. While better, they often use small convolutional kernels that miss long-range interactions or lose resolution due to pooling layers. They also struggle with the "empty space" problem—most of the grid is just empty road, which is computationally wasteful.

The authors argue that human drivers "surmise" the intentions of others through mirrors and observation. GISNet aims to emulate this behavior by establishing a formal information-sharing network among all vehicles in a scene.

Methods

GISNet is an end-to-end differentiable model composed of four main stages:

  1. Vehicle Information Embedding: The model uses Long Short-Term Memory (LSTM) networks to encode the 3-second history of every vehicle in the scene into a feature vector ($x_{traj}$).
  2. Social Tensor Construction: To preserve spatial relationships, these vectors are placed into a 13x3 grid (the "Social Tensor"). Convolutional layers then extract a social feature vector ($x_{social}$).
  3. The GIS (Graph-Based Information Sharing) Module: This is the core innovation. The vehicles are treated as nodes in a graph. A two-layer Graph Convolutional Network (GCN) allows features to propagate between nodes. Unlike a standard CNN, the GCN can explore meaningful features from irregular, non-Euclidean structures, capturing the "latent intention" of neighbors ($x_{info}$).
  4. Trajectory Generation: The model fuses the three features—the vehicle's own history, the spatial social context, and the GCN-derived interaction data. An LSTM-based decoder then predicts the $x,y$ coordinates for the next 5 seconds.

Datasets

The researchers evaluated GISNet using the NGSIM (Next Generation Simulation) dataset, specifically:

  • US-101: Southbound highway data from Los Angeles.
  • I-80: Eastbound highway data from the San Francisco Bay Area. The data was split into 8-second segments: 3 seconds of history to predict 5 seconds of future movement.

Key Insights and Ideas

  • Beyond Euclidean Space: Traffic isn't a flat image; it's a set of relationships. By using GCNs, the model can weigh the importance of a neighbor based on its "connectivity" and behavior rather than just its pixel distance in a grid.
  • Latent Intention: The GCN layers effectively allow the target vehicle to "query" its neighbors. This helps the model understand, for example, that a car in a congested lane is likely to remain stationary or attempt a lane change, even if its current velocity is zero.
  • Efficiency via Fusion Graphs: To speed up training, the authors used a "batch-wise" GCN operation, concatenating multiple individual traffic graphs into a single large block-diagonal matrix. This allows the model to process multiple scenes simultaneously without losing the unique connectivity of each instance.

Key Results

GISNet significantly outperformed all baselines, including Constant Velocity (CV), GAIL-GRU, and several variations of Social-LSTM.

  • Quantitative: Compared to MATS (Multi-Agent Tensor Fusion), GISNet achieved a 50% accuracy improvement at the 1-second horizon and maintained a 21.79% improvement at the 5-second horizon.
  • RMSE: The Root Mean Square Error (RMSE) for a 5-second prediction was reduced to 3.23 meters, compared to 4.13m for MATS and 6.27m for a vanilla LSTM.
  • Qualitative: Visualization showed that GISNet accurately predicts trajectories in diverse scenarios:
  • High-speed traffic: It matches the observed ground truth almost perfectly.
  • Congested/Crowded traffic: It correctly identifies when a vehicle is "trapped" in a lane and predicts stationary behavior or complex lane-change maneuvers that other models miss.

Limitations and Future Work

While GISNet is a leap forward, the authors acknowledge a few areas for growth:

  • Static Topology: The current model builds the graph based on the relative positions at the last observed timestamp. In reality, the "social network" of cars is fluid; a car that was a neighbor three seconds ago might be far away now.
  • Dynamic Relationships: Future work aims to explore how the information-sharing network should evolve over time, potentially using dynamic graph structures to better reflect the changing nature of highway traffic.

Paper Citation: Ziyi Zhao, Haowen Fang, Zhao Jin, Qinru Qiu. "GISNet: Graph-Based Information Sharing Network For Vehicle Trajectory Prediction." arXiv:2003.11973v1 [cs.RO] 22 Mar 2020.

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},
}

Loading related content...

Figure