PBP: Path-based Trajectory Prediction for Autonomous Driving

TL;DR

This paper proposes the Path-based prediction (PBP) approach, which predicts a discrete probability distribution over reference paths in the HD map using the path features and predicts trajectories in the path-relative Frenet frame and achieves competitive performance on the standard trajectory prediction metrics.

Abstract

Trajectory prediction plays a crucial role in the autonomous driving stack by enabling autonomous vehicles to anticipate the motion of surrounding agents. Goal-based prediction models have gained traction in recent years for addressing the multimodal nature of future trajectories. Goal-based prediction models simplify multimodal prediction by first predicting 2D goal locations of agents and then predicting trajectories conditioned on each goal. However, a single 2D goal location serves as a weak inductive bias for predicting the whole trajectory, often leading to poor map compliance, i.e., part of the trajectory going off-road or breaking traffic rules. In this paper, we improve upon goal-based prediction by proposing the Path-based prediction (PBP) approach. PBP predicts a discrete probability distribution over reference paths in the HD map using the path features and predicts trajectories in the path-relative Frenet frame. We applied the PBP trajectory decoder on top of the HiVT scene encoder and report results on the Argoverse dataset. Our experiments show that PBP achieves competitive performance on the standard trajectory prediction metrics, while significantly outperforming state-of-the-art baselines in terms of map compliance.

Figures (4)

• Figure 1: Overview of path-based prediction. Path-based prediction predicts trajectories conditioned on reference paths rather than 2D goals. We sample reference paths using the lane network from HD maps, predict a discrete distribution over the sampled paths, and predict future trajectories in the Frenet frame relative to the paths. Finally, we transform the trajectories back to the Cartesian frame relative to the target agent to obtain multimodal predictions.
• Figure 2: Model architecture: Our model consists of four key modules. The scene encoder encodes the agent history and HD map information (Section \ref{['sec:scene_enc']}). The candidate path sampler samples candidate paths for each agent from the lane graph (Section \ref{['sec:target_sampling']}). The path classifier predicts a discrete distribution over the reference paths (Section \ref{['sec:path_class']}). Finally, the trajectory regressor decodes trajectory predictions in the path-relative Frenet frame conditioned on the paths (Section \ref{['sec:decoder']}).
• Figure 3: Offroad rate.
• Figure 4: Qualitative comparison between original HiVT-64 and PBP. The first column shows the predictions from HiVT-64, and the second column shows the predictions from PBP. The blue, green, and red lines represent past history, ground-truth, and top-6 prediction trajectories, respectively.