Flexible anchor-based trajectory prediction for different types of traffic participants in autonomous driving systems

Abstract

The task of trajectory prediction is a critical component of autonomous vehicle systems. Existing trajectory prediction methodologies encounter challenges in effectively handling varied traffic participant categories and accurately forecasting long-term trajectories. In response, we introduce the Fourier Transformer Prediction (FTP) framework, which integrates the Fourier transform and a flexible anchor approach. The Fourier transform encoder adeptly captures temporal and spectral domain features inherent in trajectory data across diverse categories. The flexible anchor method employs a proposal module without anchors to generate adaptable coarse trajectories, complemented by an anchor-based module for subsequent refinement. FTP adeptly models the characteristics of multi-class participants, enhancing training stability and mitigating issues such as mode collapse. Through extensive experiments conducted on public datasets and the proposed Traffic Route Bus trajectory prediction dataset (TRB), FTP demonstrates superior performance, underscoring its efficacy across diverse traffic scenarios.