The unpredictable nature of human behaviour is a critical aspect in the domain of trajectory prediction, especially when viewed in the context of autonomous vehicles. Generational models such as Conditional Variational Autoencoders (CVAEs) or Generative Adversarial Networks (GANs) have been shown to produce state-of-the-art results, however, there is still room for improvements in areas such as the ability to generate natural, collision-free predictions.
Denoising Diffusion Probabilistic Models (DDPMs) are a recent type of generative model that has seen substantial adaptation across fields such as image generation, audio synthesis, and time series forecasting, but has yet to infiltrate trajectory prediction thoroughly. In this paper, we make further advancements in this domain, demonstrating DDPMs' ability to generate natural, scene-compliant trajectories, capable of competing with state-of-the-art methods through our proposed model, SCTP.
We address the slow inference time of DDPMs - a critical aspect of trajectory prediction when viewed in the context of autonomous vehicles - by adopting the leapfrogging technique and a point cloud representation of the map to decrease the inference speed by a factor of more than two when compared to traditional approaches. We also demonstrate that closed-loop predictions can perform as well as state-of-the-art approaches, which commonly generate the trajectories in a one-shot approach, allowing the model to evaluate the map repeatedly, contributing to its superior performance when generating collision-free trajectories.