[Paper Review] DROGON: A Trajectory Prediction Model based on Intention-Conditioned Behavior Reasoning

Chiho Choi, Srikanth Malla+2 more|arXiv (Cornell University)|Jul 31, 2019

Autonomous Vehicle Technology and Safety21 citations

TL;DR

DROGON proposes a trajectory prediction model that conditions vehicle motion on estimated intentions using relational interaction inference, intention estimation, and behavior reasoning. It achieves state-of-the-art performance on both vehicle and pedestrian trajectory prediction benchmarks by modeling goal-oriented behavior through probabilistic intention distributions, outperforming prior methods on HID, SDD, ETH, and UCY datasets.

ABSTRACT

We propose a Deep RObust Goal-Oriented trajectory prediction Network (DROGON) for accurate vehicle trajectory prediction by considering behavioral intentions of vehicles in traffic scenes. Our main insight is that the behavior (i.e., motion) of drivers can be reasoned from their high level possible goals (i.e., intention) on the road. To succeed in such behavior reasoning, we build a conditional prediction model to forecast goal-oriented trajectories with the following stages: (i) relational inference where we encode relational interactions of vehicles using the perceptual context; (ii) intention estimation to compute the probability distributions of intentional goals based on the inferred relations; and (iii) behavior reasoning where we reason about the behaviors of vehicles as trajectories conditioned on the intentions. To this end, we extend the proposed framework to the pedestrian trajectory prediction task, showing the potential applicability toward general trajectory prediction.

Motivation & Objective

• To develop a robust trajectory prediction model that accounts for driver intentions in complex traffic scenes.
• To address the lack of intention-aware modeling in existing vehicle trajectory prediction frameworks, especially in intersection scenarios.
• To create a large-scale, interactive dataset (HID) with labeled intention zones for training and evaluation.
• To generalize the framework to pedestrian trajectory prediction, validating its broader applicability.
• To improve prediction accuracy by conditioning behavior on estimated intentions, enhancing physical plausibility and multi-modality.

Proposed method

• Uses relational inference to encode interactions between vehicles and the environment based on perceptual context.
• Estimates probability distributions over potential goal zones (intentions) using inferred relational features.
• Applies conditional behavior reasoning to generate multi-modal trajectories based on estimated intentions.
• Extends the framework to pedestrians by treating all grid regions as potential intention zones, enabling generalization.
• Employs a probabilistic, goal-conditioned generation mechanism that improves robustness to intention misclassification.
• Uses a 5×5 grid partitioning of the image space to define intention categories, enabling cross-domain evaluation.

Experimental results

Research questions

• RQ1 Can a trajectory prediction model effectively estimate driver intentions from perceptual and relational context in traffic scenes?
• RQ2 How does conditioning trajectory prediction on estimated intentions improve accuracy and physical plausibility?
• RQ3 Can the proposed framework generalize from vehicle to pedestrian trajectory prediction in diverse environments?
• RQ4 What is the impact of uncertainty in intention estimation on final prediction performance?
• RQ5 How does DROGON compare to state-of-the-art models across diverse benchmark datasets?

Key findings

• DROGON-E achieves the best performance on the SDD benchmark, with FDE of 1.24 at 1.0 sec, 2.19 at 2.0 sec, 3.36 at 3.0 sec, 4.94 at 4.0 sec, and 30.90 FDE at 4.8 sec (in pixels at 1/5 resolution).
• On the ETH and UCY datasets, DROGON-E achieves an average ADE of 0.41 m and FDE of 0.65 m at 4.8 sec, outperforming S-GAN, SoPhie, S-BiGAT, PMP-NMMP, and S-STGCNN.
• The DROGON-Prob-20 variant, which uses adaptive conditioning on intention probabilities, shows improved robustness to intention misclassification.
• DROGON successfully predicts motion for static vehicles (e.g., stopped at intersections), enabling collision avoidance through intention-aware forecasting.
• The framework generalizes effectively to pedestrian motion prediction, validating its applicability beyond vehicles.
• The creation of the Honda Intersection Dataset (HID) enables evaluation of goal-oriented behavior in highly interactive, real-world intersection scenarios.