[Paper Review] Map-Adaptive Goal-Based Trajectory Prediction

Lingyao Zhang, Po-Hsun Su+3 more|arXiv (Cornell University)|Sep 9, 2020

Autonomous Vehicle Technology and Safety26 references30 citations

TL;DR

GoalNet predicts multi-modal, long-term vehicle trajectories by proposing map-derived goal paths and modeling both goal-based and motion-based modes with a graph network. It adapts to arbitrary map geometries and outperforms baselines on internal and nuScenes datasets.

ABSTRACT

We present a new method for multi-modal, long-term vehicle trajectory prediction. Our approach relies on using lane centerlines captured in rich maps of the environment to generate a set of proposed goal paths for each vehicle. Using these paths -- which are generated at run time and therefore dynamically adapt to the scene -- as spatial anchors, we predict a set of goal-based trajectories along with a categorical distribution over the goals. This approach allows us to directly model the goal-directed behavior of traffic actors, which unlocks the potential for more accurate long-term prediction. Our experimental results on both a large-scale internal driving dataset and on the public nuScenes dataset show that our model outperforms state-of-the-art approaches for vehicle trajectory prediction over a 6-second horizon. We also empirically demonstrate that our model is better able to generalize to road scenes from a completely new city than existing methods.

Motivation & Objective

• Leverage high-definition map lane centerlines to infer actor goals and drive long-horizon trajectory prediction.
• Develop a map-adaptive framework that can handle arbitrary lane topologies via a variable number of goals.
• Integrate goal-based and motion-based trajectory modes to capture both compliant and non-compliant driving behaviors.
• Employ a path-relative coordinate frame to improve generalization across road geometries and scenes.

Proposed method

• Generate a variable set of goal paths for each actor from lane centerlines within a search radius.
• Encode actor state, along-path context, and goal paths using an encoder and path-aligned rasters.
• Use a graph network with one actor node and multiple goal nodes to predict trajectories tied to each goal.
• Represent outputs in a path-relative frame to separate along-track and cross-track components.
• Model spatial and temporal multimodality with goal-based and goal-free (motion-based) trajectory modes.
• Train with a joint loss combining a classification loss over modes and a regression loss over trajectory components.

Experimental results

Research questions

• RQ1 Can map-derived goal paths provide a compact, interpretable multi-modal forecast for each actor?
• RQ2 Does a map-adaptive, set-based prediction framework improve long-horizon accuracy and generalization to new cities?
• RQ3 How does integrating goal-based and motion-based modes affect predictive coverage and error metrics?
• RQ4 What is the impact of the path-relative frame on learning and generalization across diverse road geometries?

Key findings

• GoalNet achieves state-of-the-art performance on internal and nuScenes datasets across multiple metrics.
• GoalNet-1T (1 temporal mode per spatial mode) and GoalNet-2T (2 temporal modes) show strong results, with GoalNet-2T often excelling on most metrics.
• The model uses a variable number of trajectory modes, adapting to scene geometry, reducing the need for an excessive fixed trajectory set.
• A substantial reduction in cross-track error is observed at longer horizons, highlighting advantages for long-horizon prediction.
• Inter-city generalization shows GoalNet degrades less than baselines when tested in an unseen city, demonstrating map-adaptive compositional inductive bias.
• Ablations show path rasters and path-relative projections contribute to improved ATE/CTE and overall prediction quality.