Conditional Goal-Oriented Trajectory Prediction for Interacting Vehicles.

Area of Science:

• Autonomous Driving
• Computer Vision
• Robotics

Background:

• Predicting pairwise traffic agent trajectories in interactive scenarios (e.g., cut-ins, merging) is a significant challenge for autonomous driving systems.
• Existing methods often fail to adequately model future interactions, either by treating prediction marginally or using single-axis factorization.

Purpose of the Study:

• To propose a novel double-axis factorized joint prediction pipeline, the conditional goal-oriented trajectory prediction (CGTP) framework.
• To effectively model future interactions along both agent and time axes for improved goal and trajectory prediction.

Main Methods:

• Designed a goals-of-interest network (GoINet) for hierarchical vectorized feature extraction of goal candidates.
• Developed a conditional goal prediction network (CGPNet) for multimodal goal pair prediction with a novel goal interactive loss.
• Proposed a goal-oriented trajectory rollout network (GTRNet) for scene-compliant trajectory pair prediction via timewise interactive rollouts, guided by predicted goal pairs.

Main Results:

• The proposed CGTP framework demonstrates superior performance compared to state-of-the-art (SOTA) prediction models.
• Validation was conducted on diverse datasets including the Waymo Open Motion Dataset (WOMD), Argoverse Motion Forecasting Dataset, and an in-house cut-in dataset.

Conclusions:

• The CGTP framework effectively addresses the challenge of interactive trajectory prediction in autonomous driving.
• The double-axis factorization approach, considering both agent and time interactions, leads to significant improvements in prediction accuracy and scene compliance.