Characterized Diffusion and Spatial-Temporal Interaction Network for Trajectory Prediction in Autonomous Driving

TL;DR

The paper introduces CDSTraj, a generative trajectory prediction model for autonomous driving that jointly models uncertainty and agent interactions through a Characterized Diffusion Module and a Spatio-Temporal Interaction Network. By simulating uncertain future scenes with forward diffusion and denoising via a conditional context and a multi-modal decoder, the approach yields diverse, realistic future trajectories. The method achieves state-of-the-art results on NGSIM, HighD, and MoCAD across short and long horizons, supported by thorough ablations and qualitative analyses. This work advances AD planning by enabling more reliable predictions in heterogeneous, dynamic traffic contexts.

Abstract

Trajectory prediction is a cornerstone in autonomous driving (AD), playing a critical role in enabling vehicles to navigate safely and efficiently in dynamic environments. To address this task, this paper presents a novel trajectory prediction model tailored for accuracy in the face of heterogeneous and uncertain traffic scenarios. At the heart of this model lies the Characterized Diffusion Module, an innovative module designed to simulate traffic scenarios with inherent uncertainty. This module enriches the predictive process by infusing it with detailed semantic information, thereby enhancing trajectory prediction accuracy. Complementing this, our Spatio-Temporal (ST) Interaction Module captures the nuanced effects of traffic scenarios on vehicle dynamics across both spatial and temporal dimensions with remarkable effectiveness. Demonstrated through exhaustive evaluations, our model sets a new standard in trajectory prediction, achieving state-of-the-art (SOTA) results on the Next Generation Simulation (NGSIM), Highway Drone (HighD), and Macao Connected Autonomous Driving (MoCAD) datasets across both short and extended temporal spans. This performance underscores the model's unparalleled adaptability and efficacy in navigating complex traffic scenarios, including highways, urban streets, and intersections.

Figures (4)

• Figure 1: Overview of our model for processing past states. The framework utilizes two specialized modules to accomplish trajectory prediction for the target agent: Characterized Diffusion and Spatial-Temporal Interaction Network. In situations of high uncertainty, characterized diffusion employs a noisy Gaussian function to define a confidence region for the trajectory distribution. Continuous denoising isolates confidence features for future predictions. Meanwhile, the spatial-temporal interaction network extracts features to understand spatial relations and temporal dependency.
• Figure 2: Framework of the proposed model.
• Figure 3: Overview of the Characterized Diffusion Module. The observed historical states are used as input for forward diffusion, noise is added for every steps by the Gaussian distributions with controllable means and variances. When $t = n$, the model performs reverse diffusion guided by the conditional embedding between each step to estimate the noise from the sampled noises with a Linear network, and finally output future trajectories.
• Figure 4: Visualization of our proposed model.