C2F-TP: A Coarse-to-Fine Denoising Framework for Uncertainty-Aware Trajectory Prediction

Zichen Wang, Hao Miao, Senzhang Wang, Renzhi Wang, Jianxin Wang, Jian Zhang · 2024

Read C2F-TP: A Coarse-to-Fine Denoising Framework for Uncertainty-Aware Trajectory Prediction by Zichen Wang, Hao Miao, Senzhang Wang, Renzhi Wang, Jianxin Wang, Jian Zhang (2024) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2412-13231.txt.

Introduction

Zichen Wang and five co-authors predict car paths in a 2024 paper on self-driving safety. A car that can guess where nearby cars will go can plan safer moves. True paths stay unsure, since driver aims shift and road scenes vary so widely.

Their frame, called C2F-TP, refines each guess in two passes, from coarse to fine. A first stage reads how cars sway each other across place and time. A second stage cleans the rough paths into sharp ones, with doubt kept in view. Tests show the two-pass plan tracks real car paths with strong care.

Catalog record

CREATORZichen Wang, Hao Miao, Senzhang Wang, Renzhi Wang, Jianxin Wang, Jian Zhang
YEAR2024
COLLECTIONPAPERS
FORMATTXT
FILE SIZE43 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2412-13231

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

C2F-TP: A Coarse-to-Fine Denoising Framework for Uncertainty-Aware Trajectory Prediction Zichen Wang, Hao Miao, Senzhang Wang, Renzhi Wang, Jianxin Wang, Jian Zhang Abstract Accurately predicting the trajectory of vehicles is critically important for ensuring safety and reliability in autonomous driving. Although considerable research efforts have been made recently, the inherent trajectory uncertainty caused by various factors including the dynamic driving intends and the diverse driving scenarios still poses significant challenges to accurate trajectory prediction. To address this issue, we propose C2F-TP, a coarse-to-fine denoising framework for uncertainty-aware vehicle trajectory prediction. C2F-TP features an innovative two-stage coarse-to-fine prediction process. Specifically, in the spatial-temporal interaction stage, we propose a spatial-temporal interaction module to capture the inter-vehicle interactions and learn a multimodal trajectory distribution, from which a certain number of noisy trajectories are sampled. Next, in the trajectory refinement stage, we design a conditional denoising model to reduce the uncertainty of the sampled trajectories through a step-wise denoising operation. Extensive experiments are conducted on two real datasets NGSIM and highD that are widely adopted in trajectory prediction. The result demonstrates the effectiveness of our proposal. Code — https://github.com/wangzc0422/C2F-TP Introduction Vehicle trajectory prediction aims to predict the future trajectory of the target agent based on the historical trajectories of itself and its surrounding neighbours. Accurately predicting the future trajectory of vehicles is crucial for many autonomous driving applications, including optimal driving path planning, making accurate driving decisions in dynamic environments, and enhancing driving safety ( Liu et al. 2021 ; Lun et al. 2024 ) . Figure 1: Illustration of the dynamics and temporal correlation of inter-vehicle interactions. Traditionally, statistical models are employed to predict future trajectories based on the historical trajectories of individual agents ( Hearst et al. 1998 ; Williams and Rasmussen 2006 ) . Nonetheless, they do not consider the interactions between the target agent and the surrounding agents, which deteriorates the prediction performance. To address this issue, various deep learning based models ( Alahi et al. 2016 ; Deo and Trivedi 2018 ; Salzmann et al. 2020 ; Song et al. 2020 ; Zhao et al. 2021 ) are proposed to model the spatial interactions among vehicles. However, existing deep learning based approaches fall short in modeling the uncertainty of the trajectories ( Chen et al. 2022 ; Wang et al. 2023 ) , which is common in real-world driving scenarios ( Liao et al. 2024 ) . Generally, uncertainty in autonomous driving can be broadly categorized into uncertainty in driving behavior and in driving scenarios ( Suk et al. 2024 ) . How to design an uncertainty-aware trajectory prediction model is still not fully explored and remains as an open research problem. However, achieving the uncertainty-aware trajectory prediction is non-trivial due to the following challenges. First, the inherent uncertainty in trajectories caused by various uncertainty factors in dynamic driving scenarios, make the accurate future trajectory prediction difficult ( Li et al. 2023 ) . Existing approaches mostly directly predict the coordinate points of future trajectories ( Alahi et al. 2016 ; Xie et al. 2021 ) without considering such uncertainty factors. Although some recent methods are proposed for multimodal trajectory prediction ( Deo and Trivedi 2018 ; Chen et al. 2022 ) , they still cannot effectively model the complex uncertainty. Second, it is challenging to simultaneously model the dynamics and temporal correlations of vehicle interactions. Existing approaches either only learn a fixed interaction weight between two vehicles ( Alahi et al. 2016 ; Deo and Trivedi 2018 ) or simply consider the dynamics of interactions by modeling them at each timestamp ( Wang et al. 2023 ) . For example, as shown in Figure 1 , at time t , vehicle A is driving at a constant speed, and vehicles B and C intend to accelerate to merge into the middle lane. To avoid a collision, vehicle A may slow down to keep a safe distance from vehicles B and C , and vehicle C may delay the lane-changing considering vehicle B ’s lane-changing behavior at time t+1 . Third, it is challenging to model the multimodal driving intentions. Most existing methods ( Deo and Trivedi 2018 ; Wang et al. 2023 ; Miao et al. 2024b ) perform trajectory prediction by considering multiple intentions simultaneously. However, different motion intentions have diverse impacts on the future trajectories. Simply combining them together may lead to a suboptimal performance. To address these challenges, we propose a coarse-to-fine denoising framework named C2F-TP for uncertainty-aware trajectory prediction. C2F-TP features a novel two-stage prediction process including a spatial-temporal interaction module and a refinement module. Specifically, in the spatial-temporal interaction module, C2F-TP aims to learn a multimodal future trajectory distribution based on historical trajectories. Next, a certain number of noisy trajectories are sampled from the learned distribution and fed into the refinement module. The refinement module adopts a conditional denoising model to generate stable and reliable future trajectories through stepwise denoising, greatly reducing the uncertainty of the input trajectories. In particular, in the spatial-temporal interaction module, motivated by ( Wang et al. 2023 ) we propose an interaction pooling to simultaneously capture the dynamic and temporal correlations of interactions across vehicles. To adaptively learn the impact of different motion modes on trajectory prediction, we propose a re-weighted multi-modal trajectory predictor to embed each motion mode separately. The learned embeddings are then fused adaptively by considering their diverse importance to future trajectory prediction. Moreover, we learn a multimodal future trajectory distribution based on the fused embeddings. Finally, we sample k coarse-grained trajectories from the trajectory distribution and feed them into the refinement stage. The refinement module uses historical trajectories as a condition to refine these k coarse-grained trajectories, reducing their uncertainty through stepwise denoising and improving the accuracy of trajectory prediction. The primary contributions of this paper are as follows. • We for the first time propose a novel coarse-to-fine trajectory prediction model with a two-stage generation process, which considers the complex interactions across vehicles in trajectory prediction. • To capture the complex spatial and temporal interactions, we design a spatial-temporal interaction module with a re-weighted multi-modal trajectory predictor to model the sophisticated trajectory distributions. • To contend with the inherent stochastic nature of trajectories, we design a trajectory refinement stage based on the denoising diffusion model. Related Work Trajectory Prediction. Trajectory prediction attracts increasing interest due to the increasing availability of trajectory data ( Krajewski et al. 2018 ; Deo and Trivedi 2018 ) . Recently, significant advances have been achieved in developing deep learning based methods ( Liu et al. 2021 ) for trajectory prediction. First, RNNs ( Morton, Wheeler, and Kochenderfer 2016 ; Wang, Cao, and Philip 2020 ) have been applied to predict the future movement paths of agents with historical trajectory data. However, in complex traffic environments, surrounding vehicles may significantly affect the prediction result due to the perception of interactions between agents. To address this issue, many recent studies ( Alahi et al. 2016 ; Mohamed et al. 2020 ; Chen et al. 2022 ; Wang et al. 2023 ; Miao et al. 2024a ) try to model the spatial interactions across agents by means of various emerging pooling techniques. Despite their advances, existing methods suffer from effective data uncertainty

Read the complete text