DiffTrajectory: Mitigating cumulative errors and enhancing inference efficiency in diffusion-based trajectory prediction

Chengcheng Li, Luqi Gong, Leiheng Xu, Xin Wang

Abstract: Diffusion models have made significant progress in trajectory prediction tasks but still face several critical challenges. The ordinary differential equation (ODE) solving methods used in standard diffusion models often suffer from error accumulation during multi-step iterations. Additionally, the denoising process is highly time-consuming due to the large number of computational steps, which significantly hinders inference efficiency and makes real-time applications challenging. To address these issues, we propose a diffusion-based method, DiffTrajectory, which integrates the Runge-Kutta (RK4) method, a Leap Initializer Module (LIM), and an Adaptive Dynamic Step-size Strategy (ADSS) to enhance generation accuracy and greatly optimize inference efficiency. Specifically, to tackle the problem of error accumulation, DiffTrajectory formalizes the denoising process as an ODE-solving problem and adopts the RK4 as a numerical solution. By computing multiple intermediate points at each iteration, this approach significantly reduces error accumulation. To improve the efficiency of the denoising process, DiffTrajectory introduces LIM, which leverages a pre-trained initial model to quickly generate a high-quality starting point for denoising, thereby reducing the computational burden during the initial denoising stages. Furthermore, we design the ADSS that adjusts the step size dynamically based on the results of each denoising stage, ensuring the quality of the generated results while substantially shortening inference time. Extensive experiments on the ETH/UCY and NBA datasets demonstrate that DiffTrajectory achieves substantial improvements in both accuracy and efficiency.

External IDs:doi:10.1016/j.patcog.2025.112339