Map-free vehicle trajectory prediction method based on heterogeneous graphs and dynamic scene constraints

Area of Science:

• Computer Science
• Robotics
• Artificial Intelligence

Background:

• Accurate vehicle motion prediction is vital for autonomous driving systems.
• Existing models struggle with incomplete or unavailable high-definition (HD) map data.
• Map-free prediction methods are needed to overcome these limitations.

Purpose of the Study:

• To propose a novel map-free framework for vehicle trajectory prediction, named Dynamic Scene Prediction (DynaScene-Pred).
• To compensate for the lack of explicit map data by extracting implicit road topology.
• To generate diverse and physically plausible multi-modal trajectories.

Main Methods:

• Extracted implicit road topology priors using OPTICS clustering and Bidirectional Long Short-Term Memory (Bi-LSTM) networks to create virtual lane features.
• Constructed a heterogeneous graph with vehicle and virtual lane nodes.
• Employed graph convolutional networks (GCN) with multi-dimensional attention (spatial, lane, temporal) and a Conditional Variational Autoencoder (CVAE) decoder.

Main Results:

• DynaScene-Pred demonstrated competitive performance against existing map-free baselines on the Argoverse benchmark.
• Achieved relative improvements of 5.7% and 1.8% in [Formula: see text] and [Formula: see text] respectively, compared to the TR-Pred model on the validation set.
• Effectively captured complex vehicle-vehicle interactions and vehicle-lane spatial constraints.

Conclusions:

• The proposed DynaScene-Pred framework offers a robust solution for vehicle trajectory prediction without relying on HD maps.
• Implicit road topology extraction and advanced graph neural networks enable effective scene understanding and prediction.
• The CVAE decoder ensures the generation of diverse and realistic future trajectories, crucial for safe autonomous navigation.