NEST: A Neuromodulated Small-world Hypergraph Trajectory Prediction Model for Autonomous Driving

TL;DR

Trajectory prediction for autonomous driving is challenged by non-linear, uncertain interactions in dense traffic. NEST introduces a Neuromodulated Small-world Hypergraph that combines a small-world network with hypergraph learning, enhanced by a neuromodulator to adapt to traffic conditions, and attaches a Context Fusion module with a multi-modal Laplace-based predictor. Given historical data $X$ and maps $M$, NEST outputs $K$ trajectory hypotheses $\textbf{Y}=[\textbf{Y}_1,\ldots,\textbf{Y}_K]$, each with $\textbf{Y}_i^t=[x_i^t,y_i^t,b_{i,x}^t,b_{i,y}^t]$ and mode probability $P_i$, achieving state-of-the-art results on nuScenes, MoCAD, and HighD while maintaining real-time inference. The approach demonstrates strong generalization, efficiency, and temporal foresight, offering a robust solution for reliable trajectory prediction in complex traffic environments.

Abstract

Accurate trajectory prediction is essential for the safety and efficiency of autonomous driving. Traditional models often struggle with real-time processing, capturing non-linearity and uncertainty in traffic environments, efficiency in dense traffic, and modeling temporal dynamics of interactions. We introduce NEST (Neuromodulated Small-world Hypergraph Trajectory Prediction), a novel framework that integrates Small-world Networks and hypergraphs for superior interaction modeling and prediction accuracy. This integration enables the capture of both local and extended vehicle interactions, while the Neuromodulator component adapts dynamically to changing traffic conditions. We validate the NEST model on several real-world datasets, including nuScenes, MoCAD, and HighD. The results consistently demonstrate that NEST outperforms existing methods in various traffic scenarios, showcasing its exceptional generalization capability, efficiency, and temporal foresight. Our comprehensive evaluation illustrates that NEST significantly improves the reliability and operational efficiency of autonomous driving systems, making it a robust solution for trajectory prediction in complex traffic environments.

Figures (3)

• Figure 1: Comparison of interaction modeling methods. (a) Traditional graph-based approach requires 24 edges to represent agent interactions. (b) Hypergraph method reduces this complexity to 4 hyperedges by grouping interactions. (c) Our Small-world Hypergraph further refines this by incorporating latent connections, effectively capturing long-range interactions typical of traffic scenarios. The use of hyperedges and latent links allows for a more efficient and comprehensive representation of agent interactions.
• Figure 2: The overview of the proposed NEST model. Panel (a) illustrates the overall framework of NEST, which takes agent historical data and high-definition (HD) maps as inputs. The NEST model processes these inputs with four key modules: Hypergraph Forming, Hypergraph Pooling, Context Fusion, and a Multi-modal Predictor. These modules work in tandem to predict the target agent's intention and future trajectory. Panels (b) and (c) provide a detailed breakdown of the components involved in forming the interaction hypergraph, specifically focusing on the Neuromodulator and the Small-world Network.
• Figure 3: Qualitative comparison of our NEST model with various models. Panels (a) and (b) visualize the most probable prediction by each model, whereas panel (c) visualizes predictions across ten modalities. Panel (a) illustrates the scenario where a distant leading vehicle accelerates. In this scenario, our NEST model maintains a trajectory prediction that aligns closely with the ground truth. In contrast, other models are misled by the stationary state of surrounding vehicles, resulting in slower predictions. Panel (b) depicts a scenario where a distant leading vehicle remains stationary, significantly influencing the expected speed of the target agent. Again, only our NEST model predicts a trajectory that mirrors the ground truth closely, while other models predict overly fast trajectories. Panel (c) distinctly showcases the efficacy of hypergraph-based models (NEST and Method B) in predicting the driver's true intent through their prediction of a complex U-turn maneuver.