📝 SummarXiv

Abstract

In autonomous driving, trajectory prediction is essential for ensuring safe and efficient navigation. To improve prediction accuracy, recent approaches often rely on pre-built high-definition (HD) maps or real-time local map construction modules to incorporate static environmental information. However, pre-built HD maps are limited to specific regions and cannot adapt to transient changes. In addition, local map construction modules, which recognize only predefined elements, may fail to capture critical scene details or introduce errors that degrade prediction performance. To overcome these limitations, we propose Bird's-Eye View Trajectory Prediction (BEVTraj), a novel trajectory prediction framework that operates directly in the bird's-eye view (BEV) space utilizing real-time sensor data without relying on any pre-built maps. The BEVTraj leverages deformable attention to efficiently extract relevant context from dense BEV features. Furthermore, we introduce a Sparse Goal Candidate Proposal (SGCP) module, which enables full end-to-end prediction without requiring any post-processing steps. Extensive experiments demonstrate that the BEVTraj achieves performance comparable to state-of-the-art HD map-based models while offering greater flexibility by eliminating the dependency on pre-built maps. The source code is available at https://github.com/Kongminsang/bevtraj.