國立虎尾科技大學 |

A Learning-Based Integrated Framework of Motion Prediction and Planning for Connected and Automated Vehicles : = Towards Interaction, Multi-Modality, and Relational Reasoning.

Record Type:	Language materials, manuscript : Monograph/item
Title/Author:	A Learning-Based Integrated Framework of Motion Prediction and Planning for Connected and Automated Vehicles :/
Reminder of title:	Towards Interaction, Multi-Modality, and Relational Reasoning.
Author:	Wu, Keshu.
Description:	1 online resource (149 pages)
Notes:	Source: Dissertations Abstracts International, Volume: 85-12, Section: B.
Contained By:	Dissertations Abstracts International85-12B.
Subject:	Transportation. -
Online resource:	click for full text (PQDT)
ISBN:	9798382837109

A Learning-Based Integrated Framework of Motion Prediction and Planning for Connected and Automated Vehicles : = Towards Interaction, Multi-Modality, and Relational Reasoning.
Wu, Keshu.

A Learning-Based Integrated Framework of Motion Prediction and Planning for Connected and Automated Vehicles :Towards Interaction, Multi-Modality, and Relational Reasoning. - 1 online resource (149 pages)

Source: Dissertations Abstracts International, Volume: 85-12, Section: B.

Thesis (Ph.D.)--The University of Wisconsin - Madison, 2024.

Includes bibliographical references

Predicting vehicle trajectories and ensuring safe and efficient trajectory planning are critical for the operational efficiency and safety of automated vehicles, especially on congested multi-lane highways. In these dynamic environments, a vehicle's movement is influenced by its historical behaviors and interactions with surrounding vehicles. These complex interactions result from unpredictable motion patterns, leading to diverse modalities of driving behaviors that necessitate thorough investigation. Additionally, in multi-agent systems, dynamic interactions among agents often display cooperative and competitive behaviors. Such group-wise interactions, though common, are rarely modeled. Traditional methods, while effective in capturing pair-wise interactions, fail to represent the collective influence of groups on each other's behaviors in real-world traffic scenarios. Therefore, modeling the group-wise interactions of multi-modal driving behaviors among multiple agents is essential. In dense traffic conditions, vehicles frequently change lanes, accelerate, decelerate, and engage in complex interactions with other agents. These interactions often involve multiple possible longitudinal and lateral behaviors of various entities influencing each other simultaneously, which cannot be fully captured by considering only pair-wise relationships. Furthermore, the stochastic nature of human behavior adds complexity, requiring models that handle the uncertainty and variability in agent behaviors for safe and efficient driving. Thus, a critical challenge lies in representing and reasoning about the diverse interactions among agents and their multiple possible behaviors to achieve socially inspired automated driving.This dissertation introduces the Graph-based Interaction-aware Multi-modal Trajectory Prediction (GIMTP) framework, designed to probabilistically predict future vehicle trajectories by effectively capturing these interactions. Within this framework, vehicle motions are conceptualized as nodes in a time-varying graph, and traffic interactions are represented by a dynamic adjacency matrix. To comprehensively capture both spatial and temporal dependencies embedded in this dynamic adjacency matrix, the methodology employs the Diffusion Graph Convolutional Network (DGCN), providing a graph embedding of both historical and future states. Additionally, a driving intention-specific feature fusion is implemented, enabling the adaptive integration of historical and future embeddings for enhanced intention recognition and trajectory prediction. This model offers two-dimensional predictions for each mode of longitudinal and lateral driving behaviors and provides probabilistic future paths with corresponding probabilities, addressing the challenges of complex vehicle interactions and multi-modality of driving behaviors. To further facilitate interaction-aware multi-modal motion prediction for multi-agent systems, GIMTP is enhanced to Graph-based Interaction-aware Reliable Anticipative Feasible Future Estimator (GIRAFFE), which offers multi-modal predictions by considering the behaviors of multiple vehicles.Building upon the robust GIRAFFE framework, this dissertation further integrates the Relational Hypergraph Interaction-informed Neural mOtion generator and planner (RHINO), a proposed model for motion planning that revolutionizes interaction modeling and relational reasoning for trajectory prediction and planning with multiscale hypergraph representations. RHINO distinguishes itself by surpassing previous methods that primarily consider pair-wise interactions with limited relational insight. It introduces a multiscale hypergraph neural network designed to capture intricate dynamics involving both pair-wise and group-wise interactions across multiple scales. RHINO's multiscale hypergraph is engineered to be trainable, enabling the system to discern more complex interaction patterns within traffic, such as varying group sizes and the nuances of collective behaviors. For interaction representation learning, RHINO adopts a three-element format that facilitates end-to-end learning. This innovative approach allows for explicit reasoning of relational factors, including interaction strength and category, which are crucial for accurate and socially aware motion planning. Furthermore, RHINO is integrated into both a Conditional Variational Autoencoder (CVAE)-based prediction system and enhanced state-of-the-art prediction frameworks to yield socially plausible trajectories grounded in relational reasoning. The efficacy of RHINO in understanding group behavior and discerning interaction dynamics is substantiated through synthetic physics simulations, reflecting its capability to capture group behaviors and reason about the strength and category of interactions. The effectiveness of this motion planning system is validated through extensive experiments on two real-world trajectory prediction datasets. This integrated framework of motion prediction and planning, adopting the GIRAFFE framework and RHINO framework, positions it as a powerful tool in advancing the safety and efficiency of automated vehicle operations, especially in the complex and unpredictable environment of multi-lane highways.

Electronic reproduction.
Ann Arbor, Mich. :
ProQuest,
2024

Mode of access: World Wide Web

ISBN: 9798382837109Subjects--Topical Terms:

558117
Transportation.
Subjects--Index Terms:

HypergraphIndex Terms--Genre/Form:

554714
Electronic books.

A Learning-Based Integrated Framework of Motion Prediction and Planning for Connected and Automated Vehicles : = Towards Interaction, Multi-Modality, and Relational Reasoning.
LDR:06813ntm a22004097 4500 001 1149755
005 20241022112631.5
006 m o d
007 cr bn ---uuuuu
008 250605s2024 xx obm 000 0 eng d
020 $a 9798382837109
035 $a (MiAaPQ)AAI31332013
035 $a AAI31332013
040 $a MiAaPQ $b eng $c MiAaPQ $d NTU
100 1 $a Wu, Keshu. $3 1476089
245 1 2 $a A Learning-Based Integrated Framework of Motion Prediction and Planning for Connected and Automated Vehicles : $b Towards Interaction, Multi-Modality, and Relational Reasoning.
264 0 $c 2024
300 $a 1 online resource (149 pages)
336 $a text $b txt $2 rdacontent
337 $a computer $b c $2 rdamedia
338 $a online resource $b cr $2 rdacarrier
500 $a Source: Dissertations Abstracts International, Volume: 85-12, Section: B.
500 $a Advisor: Ran, Bin.
502 $a Thesis (Ph.D.)--The University of Wisconsin - Madison, 2024.
504 $a Includes bibliographical references
520 $a Predicting vehicle trajectories and ensuring safe and efficient trajectory planning are critical for the operational efficiency and safety of automated vehicles, especially on congested multi-lane highways. In these dynamic environments, a vehicle's movement is influenced by its historical behaviors and interactions with surrounding vehicles. These complex interactions result from unpredictable motion patterns, leading to diverse modalities of driving behaviors that necessitate thorough investigation. Additionally, in multi-agent systems, dynamic interactions among agents often display cooperative and competitive behaviors. Such group-wise interactions, though common, are rarely modeled. Traditional methods, while effective in capturing pair-wise interactions, fail to represent the collective influence of groups on each other's behaviors in real-world traffic scenarios. Therefore, modeling the group-wise interactions of multi-modal driving behaviors among multiple agents is essential. In dense traffic conditions, vehicles frequently change lanes, accelerate, decelerate, and engage in complex interactions with other agents. These interactions often involve multiple possible longitudinal and lateral behaviors of various entities influencing each other simultaneously, which cannot be fully captured by considering only pair-wise relationships. Furthermore, the stochastic nature of human behavior adds complexity, requiring models that handle the uncertainty and variability in agent behaviors for safe and efficient driving. Thus, a critical challenge lies in representing and reasoning about the diverse interactions among agents and their multiple possible behaviors to achieve socially inspired automated driving.This dissertation introduces the Graph-based Interaction-aware Multi-modal Trajectory Prediction (GIMTP) framework, designed to probabilistically predict future vehicle trajectories by effectively capturing these interactions. Within this framework, vehicle motions are conceptualized as nodes in a time-varying graph, and traffic interactions are represented by a dynamic adjacency matrix. To comprehensively capture both spatial and temporal dependencies embedded in this dynamic adjacency matrix, the methodology employs the Diffusion Graph Convolutional Network (DGCN), providing a graph embedding of both historical and future states. Additionally, a driving intention-specific feature fusion is implemented, enabling the adaptive integration of historical and future embeddings for enhanced intention recognition and trajectory prediction. This model offers two-dimensional predictions for each mode of longitudinal and lateral driving behaviors and provides probabilistic future paths with corresponding probabilities, addressing the challenges of complex vehicle interactions and multi-modality of driving behaviors. To further facilitate interaction-aware multi-modal motion prediction for multi-agent systems, GIMTP is enhanced to Graph-based Interaction-aware Reliable Anticipative Feasible Future Estimator (GIRAFFE), which offers multi-modal predictions by considering the behaviors of multiple vehicles.Building upon the robust GIRAFFE framework, this dissertation further integrates the Relational Hypergraph Interaction-informed Neural mOtion generator and planner (RHINO), a proposed model for motion planning that revolutionizes interaction modeling and relational reasoning for trajectory prediction and planning with multiscale hypergraph representations. RHINO distinguishes itself by surpassing previous methods that primarily consider pair-wise interactions with limited relational insight. It introduces a multiscale hypergraph neural network designed to capture intricate dynamics involving both pair-wise and group-wise interactions across multiple scales. RHINO's multiscale hypergraph is engineered to be trainable, enabling the system to discern more complex interaction patterns within traffic, such as varying group sizes and the nuances of collective behaviors. For interaction representation learning, RHINO adopts a three-element format that facilitates end-to-end learning. This innovative approach allows for explicit reasoning of relational factors, including interaction strength and category, which are crucial for accurate and socially aware motion planning. Furthermore, RHINO is integrated into both a Conditional Variational Autoencoder (CVAE)-based prediction system and enhanced state-of-the-art prediction frameworks to yield socially plausible trajectories grounded in relational reasoning. The efficacy of RHINO in understanding group behavior and discerning interaction dynamics is substantiated through synthetic physics simulations, reflecting its capability to capture group behaviors and reason about the strength and category of interactions. The effectiveness of this motion planning system is validated through extensive experiments on two real-world trajectory prediction datasets. This integrated framework of motion prediction and planning, adopting the GIRAFFE framework and RHINO framework, positions it as a powerful tool in advancing the safety and efficiency of automated vehicle operations, especially in the complex and unpredictable environment of multi-lane highways.
533 $a Electronic reproduction. $b Ann Arbor, Mich. : $c ProQuest, $d 2024
538 $a Mode of access: World Wide Web
650 4 $a Transportation. $3 558117
650 4 $a Environmental engineering. $3 557376
653 $a Hypergraph
653 $a Interaction representation
653 $a Learning-based motion planning
653 $a Motion prediction
653 $a Multi-modal prediction
653 $a Relational reasoning
655 7 $a Electronic books. $2 local $3 554714
690 $a 0709
690 $a 0543
690 $a 0775
710 2 $a ProQuest Information and Learning Co. $3 1178819
710 2 $a The University of Wisconsin - Madison. $b Civil & Environmental Engr. $3 1182853
773 0 $t Dissertations Abstracts International $g 85-12B.
856 4 0 $u http://pqdd.sinica.edu.tw/twdaoapp/servlet/advanced?query=31332013 $z click for full text (PQDT)