Pedestrian Trajectory Prediction Based on Improved Diffusion with Fourier Embeddings

Boyang Fu, Jiashu Liao, Yixuan Yang, Victor Sanchez

Abstract: Predicting stochastic pedestrian trajectories is a complex task, requiring the integration of contextual information and the inherent uncertainty of human movement. Conventional generative models attempt to capture this uncertainty by mapping randomness to a latent space, producing a multimodal distribution of potential trajectories. These models, however, often fall short in complex scenarios with highly uncertain trajectories due to inadequate temporal dependency modeling. To address this shortcoming, we propose a framework that uses an improved conditional diffusion model that significantly enhances stochastic trajectory prediction. By conditioning on past trajectory data, the model iteratively adds Gaussian noise and employs a reverse generative process to output a diverse set of future trajectories. A novel denoising component merges noised predictions with historical data through a feature extractor, leveraging cross-attention mechanisms to intertwine past and future trajectories effectively. Furthermore, we enrich the framework’s temporal analysis with Fourier embeddings, improving its time-series predictive power. Rigorous benchmarking on the ETH, UCY, and SDD datasets confirms that our framework outperforms several state-of-the-art methods in generating accurate future trajectories.

External IDs:dblp:conf/icpr/FuLYS24