Multimodal Trajectory Prediction Coupled With Ego Vehicle Motion Trend Under Target Anchor-Driven Graph Attention Network

This is the authors' abstract. We don't add key points for this paper.

Trajectory prediction is a crucial technology to ensure the safe driving of intelligent vehicles on complex urban roads. Early works mainly focused on designing complicated architectures in deep learning-based prediction models. However, existing trajectory prediction methods still have three main shortcomings: 1) insufficient heterogeneous interactions between vehicles and environmental factors, such as maps; 2) inadequate consideration of the coupling relationship between prediction and planning modules; and 3) insubstantial purposive cognition leads to the limitation of future driving behaviors. To address these challenges, this article proposes a multimodal trajectory prediction model driven by target anchors using a graph attention network coupled with the motion trend of the ego vehicle (EV). The model first constructs a multilevel dynamic scene graph and a multigranularity static scene graph to describe the heterogeneous interactions between vehicles and the map accurately. Second, the model can consider the interaction between the EV’s future trajectory and the predicted information of the target vehicle (TV). Additionally, the model represents driving intent as target anchors, which can convey rich details on future distributions, reducing the complexity of the intent space. Experiments were conducted to evaluate the prediction accuracy on the Argoverse 1 dataset and its generalization capabilities on the CARLA+ROS2 simulator. The experimental results demonstrated that the proposed method can predict the trajectory of traffic participants much more accurately than the state-of-the-art methods and exhibit excellent generalization performance under different environmental conditions.