Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Boundary-Guided Trajectory Prediction for Road Aware and Physically Feasible Autonomous Driving

Ahmed Abouelazm, Mianzhi Liu, Christian Hubschneider, Yin Wu, Daniel Slieter, J. Marius Zöllner

TL;DR

The proposed approach has a superior generalization to less prevalent maneuvers and unseen out-of-distribution scenarios, reducing the off-road rate under adversarial attacks from 66 % to just 1 % and the effectiveness of the approach in generating feasible and robust predictions is highlighted.

Abstract

Accurate prediction of surrounding road users' trajectories is essential for safe and efficient autonomous driving. While deep learning models have improved performance, challenges remain in preventing off-road predictions and ensuring kinematic feasibility. Existing methods incorporate road-awareness modules and enforce kinematic constraints but lack plausibility guarantees and often introduce trade-offs in complexity and flexibility. This paper proposes a novel framework that formulates trajectory prediction as a constrained regression guided by permissible driving directions and their boundaries. Using the agent's current state and an HD map, our approach defines the valid boundaries and ensures on-road predictions by training the network to learn superimposed paths between left and right boundary polylines. To guarantee feasibility, the model predicts acceleration profiles that determine the vehicle's travel distance along these paths while adhering to kinematic constraints. We evaluate our approach on the Argoverse-2 dataset against the HPTR baseline. Our approach shows a slight decrease in benchmark metrics compared to HPTR but notably improves final displacement error and eliminates infeasible trajectories. Moreover, the proposed approach has superior generalization to less prevalent maneuvers and unseen out-of-distribution scenarios, reducing the off-road rate under adversarial attacks from 66% to just 1%. These results highlight the effectiveness of our approach in generating feasible and robust predictions.

Boundary-Guided Trajectory Prediction for Road Aware and Physically Feasible Autonomous Driving

TL;DR

The proposed approach has a superior generalization to less prevalent maneuvers and unseen out-of-distribution scenarios, reducing the off-road rate under adversarial attacks from 66 % to just 1 % and the effectiveness of the approach in generating feasible and robust predictions is highlighted.

Abstract

Accurate prediction of surrounding road users' trajectories is essential for safe and efficient autonomous driving. While deep learning models have improved performance, challenges remain in preventing off-road predictions and ensuring kinematic feasibility. Existing methods incorporate road-awareness modules and enforce kinematic constraints but lack plausibility guarantees and often introduce trade-offs in complexity and flexibility. This paper proposes a novel framework that formulates trajectory prediction as a constrained regression guided by permissible driving directions and their boundaries. Using the agent's current state and an HD map, our approach defines the valid boundaries and ensures on-road predictions by training the network to learn superimposed paths between left and right boundary polylines. To guarantee feasibility, the model predicts acceleration profiles that determine the vehicle's travel distance along these paths while adhering to kinematic constraints. We evaluate our approach on the Argoverse-2 dataset against the HPTR baseline. Our approach shows a slight decrease in benchmark metrics compared to HPTR but notably improves final displacement error and eliminates infeasible trajectories. Moreover, the proposed approach has superior generalization to less prevalent maneuvers and unseen out-of-distribution scenarios, reducing the off-road rate under adversarial attacks from 66% to just 1%. These results highlight the effectiveness of our approach in generating feasible and robust predictions.
Paper Structure (18 sections, 2 equations, 5 figures, 4 tables)

This paper contains 18 sections, 2 equations, 5 figures, 4 tables.

Table of Contents

  1. Introduction
  2. related work
  3. Physical Feasibility
  4. Road Awareness
  5. Methodology
  6. Boundary Set Generation
  7. Network Architecture
  8. Boundary Decoder
  9. Prediction Head
  10. Pure Pursuit Layer
  11. Classification Head
  12. experimental Setup
  13. Evaluation
  14. Boundary Set Generation
  15. Prediction Accuracy
  16. ...and 3 more sections

Figures (5)

  • Figure 1: Extracted boundaries for forward (blue) and right (green) driving directions. Reachable lanes are identified based on the agent's state, and equidistant points are sampled along the left and right boundaries. The network learns superposition weights between each driving direction's left and right boundaries to generate a superposition path. Furthermore, the network estimates an acceleration profile to transform the superposition path into a trajectory. The right driving direction predictions are omitted for simplicity.
  • Figure 2: We extend the scene representation by incorporating the generated boundary set. Each boundary consists of a pair of the leftmost and rightmost boundary lanes for a permissible driving direction, sampled at equal distance and represented as a polyline. Scene elements are encoded independently using a polyline encoder, and interactions between them are captured by a transformer-based encoder. The focal agent embedding is then concatenated with each boundary embedding and passed through a boundary decoder to generate mode-specific embeddings. A prediction head subsequently processes these embeddings to compute superposition weights and acceleration profiles for each embedding. Finally, the Pure Pursuit algorithm converts these outputs into feasible trajectories.
  • Figure 3: Our boundary generation algorithm represents the HD map as a graph and identifies the nearest start lanes based on the agent's current state. It then performs reachability analysis to determine goal lanes, including the leftmost and rightmost goal lanes in each driving direction. Next, a modified depth-first search algorithm is used to extract the leftmost and rightmost lanes traversed from the start to the goal. Finally, the boundary lanes are smoothed and uniformly sampled at equidistant intervals.
  • Figure 4: Examples of boundary sets with varying complexity demonstrate the adaptability of the proposed algorithm. These examples range from simple scenarios with a single boundary to more complex ones with up to six boundaries, showcasing the generation algorithm's ability to handle diverse road layouts and driving directions effectively.
  • Figure 5: HPTR and our model predictions under scene attacks show that HPTR struggles with road topology perturbations, while our approach robustly predicts feasible, on-road trajectories. The focal agent's history is in red, and the ground truth is in yellow.