Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

GRANP: A Graph Recurrent Attentive Neural Process Model for Vehicle Trajectory Prediction

Yuhao Luo, Kehua Chen, Meixin Zhu

TL;DR

This paper addresses autonomous-vehicle trajectory prediction with explicit uncertainty quantification. It introduces GRANP, a Graph Recurrent Attentive Neural Process that combines a deterministic and latent path in a RANP encoder, Graph Attention Networks for social interactions, and a Gaussian-output decoder trained via an ELBO objective. On the highD dataset, GRANP achieves state-of-the-art RMSE and substantially better NLL compared to baselines, while providing interpretable uncertainty and attention visualizations. The approach offers a scalable, uncertainty-aware framework for predicting vehicle trajectories in complex traffic scenarios, with practical implications for safer and more efficient autonomous driving. $p(Y_T|X_T,X_C,Y_C)$ is modeled with latent variables and cross-attention, and the encoder-decoder structure enables direct, tractable uncertainty quantification without heavy sampling.

Abstract

As a vital component in autonomous driving, accurate trajectory prediction effectively prevents traffic accidents and improves driving efficiency. To capture complex spatial-temporal dynamics and social interactions, recent studies developed models based on advanced deep-learning methods. On the other hand, recent studies have explored the use of deep generative models to further account for trajectory uncertainties. However, the current approaches demonstrating indeterminacy involve inefficient and time-consuming practices such as sampling from trained models. To fill this gap, we proposed a novel model named Graph Recurrent Attentive Neural Process (GRANP) for vehicle trajectory prediction while efficiently quantifying prediction uncertainty. In particular, GRANP contains an encoder with deterministic and latent paths, and a decoder for prediction. The encoder, including stacked Graph Attention Networks, LSTM and 1D convolutional layers, is employed to extract spatial-temporal relationships. The decoder is used to learn a latent distribution and thus quantify prediction uncertainty. To reveal the effectiveness of our model, we evaluate the performance of GRANP on the highD dataset. Extensive experiments show that GRANP achieves state-of-the-art results and can efficiently quantify uncertainties. Additionally, we undertake an intuitive case study that showcases the interpretability of the proposed approach. The code is available at https://github.com/joy-driven/GRANP.

GRANP: A Graph Recurrent Attentive Neural Process Model for Vehicle Trajectory Prediction

TL;DR

This paper addresses autonomous-vehicle trajectory prediction with explicit uncertainty quantification. It introduces GRANP, a Graph Recurrent Attentive Neural Process that combines a deterministic and latent path in a RANP encoder, Graph Attention Networks for social interactions, and a Gaussian-output decoder trained via an ELBO objective. On the highD dataset, GRANP achieves state-of-the-art RMSE and substantially better NLL compared to baselines, while providing interpretable uncertainty and attention visualizations. The approach offers a scalable, uncertainty-aware framework for predicting vehicle trajectories in complex traffic scenarios, with practical implications for safer and more efficient autonomous driving. is modeled with latent variables and cross-attention, and the encoder-decoder structure enables direct, tractable uncertainty quantification without heavy sampling.

Abstract

As a vital component in autonomous driving, accurate trajectory prediction effectively prevents traffic accidents and improves driving efficiency. To capture complex spatial-temporal dynamics and social interactions, recent studies developed models based on advanced deep-learning methods. On the other hand, recent studies have explored the use of deep generative models to further account for trajectory uncertainties. However, the current approaches demonstrating indeterminacy involve inefficient and time-consuming practices such as sampling from trained models. To fill this gap, we proposed a novel model named Graph Recurrent Attentive Neural Process (GRANP) for vehicle trajectory prediction while efficiently quantifying prediction uncertainty. In particular, GRANP contains an encoder with deterministic and latent paths, and a decoder for prediction. The encoder, including stacked Graph Attention Networks, LSTM and 1D convolutional layers, is employed to extract spatial-temporal relationships. The decoder is used to learn a latent distribution and thus quantify prediction uncertainty. To reveal the effectiveness of our model, we evaluate the performance of GRANP on the highD dataset. Extensive experiments show that GRANP achieves state-of-the-art results and can efficiently quantify uncertainties. Additionally, we undertake an intuitive case study that showcases the interpretability of the proposed approach. The code is available at https://github.com/joy-driven/GRANP.
Paper Structure (19 sections, 8 equations, 4 figures, 1 table)

This paper contains 19 sections, 8 equations, 4 figures, 1 table.

Table of Contents

  1. INTRODUCTION
  2. RELATED WORK
  3. Physics-based Models
  4. Machine Learning-based Models
  5. Deep Learning-based Models
  6. METHODOLOGY
  7. Problem Definition
  8. Recurrent Attentive Neural Process (RANP)
  9. Graph Attention Networks (GAT)
  10. Graph construction
  11. GAT Structure
  12. EXPERIMENT
  13. Datasets
  14. Metrics and Settings
  15. Baseline Methods
  16. ...and 4 more sections

Figures (4)

  • Figure 1: Overall framework of GRANP. GRANP primarily follows the structure of RANPs, including an encoder and a decoder with deterministic and latent paths. More detail, the Conv-MLP module contains a triple 1D convolutional layers and 4 full connected layers, and GAT module consists of two graph attention layers.
  • Figure 2: Examples of trajectory prediction of the ego vehicle under (a) going straight and (b) lane-changing scenarios. $x$-axis, $y$-axis, $z$-axis represent the lateral coordinate, longitudinal coordinate, and the time stamp, respectively. The gray, red, and yellow lines represent historical, prediction, and true trajectories, respectively. Additionally, the pink surface represents the uncertainty. In this work, we use 95% confidence intervals.
  • Figure 3: Examples of trajectory prediction of the ego vehicle under (a) going straight and (b) lane-changing scenarios. Red and yellow cells indicate ego and neighboring vehicles respectively. The three most important attention weights are visualized.
  • Figure 4: RMSE and NLL results under different parameters. (a) and (b) represents the results with different numbers of attention heads as well as (c) and (d) shows the results with different numbers of hidden dimensions.