Graph-based Trajectory Prediction with Cooperative Information

TL;DR

The paper addresses autonomous driving trajectory prediction in settings where cooperative information from connected actors may be available. It introduces a graph-based neural network that fuses historic trajectories, map data, and cooperative inputs (paths or trajectories) through four heterogeneous graphs and fusion modules inspired by LaneGCN. A training scheme randomly samples cooperative data availability and augments trajectories to reflect real-world variability, enabling robust learning even when cooperative data is incomplete or noisy. Experiments on the Argoverse dataset demonstrate substantial improvements when cooperative data are present and reveal robustness to imperfect cooperative inputs, highlighting practical applicability in real-world automated driving scenarios.

Abstract

For automated driving, predicting the future trajectories of other road users in complex traffic situations is a hard problem. Modern neural networks use the past trajectories of traffic participants as well as map data to gather hints about the possible driver intention and likely maneuvers. With increasing connectivity between cars and other traffic actors, cooperative information is another source of data that can be used as inputs for trajectory prediction algorithms. Connected actors might transmit their intended path or even complete planned trajectories to other actors, which simplifies the prediction problem due to the imposed constraints. In this work, we outline the benefits of using this source of data for trajectory prediction and propose a graph-based neural network architecture that can leverage this additional data. We show that the network performance increases substantially if cooperative data is present. Also, our proposed training scheme improves the network's performance even for cases where no cooperative information is available. We also show that the network can deal with inaccurate cooperative data, which allows it to be used in real automated driving environments.

Figures (5)

• Figure 1: (a) shows an example scene with cooperative information. Actor 1 transmits a path (gold line) along which it wants to move. Actor 2 transmits a complete trajectory (rainbow colors indicate time information), which indicates that it will accelerate and merge onto Actor 1's lane. When predicting Actor 1, the prediction problem is constrained by its path (no hypothesis for turning right is necessary) and the movement of Actor 2 (velocity has to be adjusted). Actor 3 transmits no cooperative information and has to be predicted conventionally. For the corresponding lane graph in (b), lane nodes and predecessor/successor connections are in blue, left/right connections are dashed orange. Self- and dilated connections are omitted for clarity. (c), (d), (e) show the corresponding actor graph, path graph and trajectory graph, respectively. (f) and (g) show two of the novel fusion modules. The T2L module (f) updates lane nodes with data from nearby trajectory graph nodes, whereas L2P (g) updates path nodes with data from nearby lane nodes.
• Figure 2: Overview of the network architecture. Input data is colored in light orange, trainable network layers are colored blue, network outputs are red, and other processing modules are in green. The network is composed of a temporal convolutional network (TCN) and a graph creation step, after which a series of graph convolution modules is applied, before a multi-layer perceptron (MLP) calculates the final output from the resulting actor node features.
• Figure 3: Error when predicting the AOI's trajectory when increasing the amount of cooperative data of other actors ($\theta_\text{gt}$) when $\beta=1$ and $\theta_\text{AOI} = \times$. $\theta_\text{type}{} = 0%$ (blue) means that cooperative data stems from transmitted path information, while $\theta_\text{type} = 100%$ means that it stems from transmitted planned trajectories.
• Figure 4: Prediction error for the AOI at $\theta_\text{gt} = 0%$ and $\theta_\text{AOI} = \checkmark$ at varying values of $\beta$. Prediction error at $\theta_\text{AOI} = \times$ is shown as the reference.
• Figure 5: Examples where cooperative information improves the prediction of the AOI's trajectory. Circles denote positions of actors, where the AOI is red and other actors are cyan. The AOI's ground truth trajectory is shown as a red dashed line. Best predicted hypotheses with and without cooperative information are in green and magenta, respectively. Trajectory end points are denoted with a star in the corresponding color. Lane centerlines are in grey.