【作者】 杨玲；

【导师】 李可；

【作者基本信息】 西南交通大学 ， 计算机技术（专业学位）， 2024， 硕士

【摘要】 随着人工智能和5G技术的进步,智能联网汽车规模的持续增长,使得智能交通系统(Intelligent Transportation System,ITS)受到学术界和工业界的广泛关注。ITS是一种综合应用信息、通信和控制的先进交通管理系统,其目标是提高交通系统的效率和安全性,通过实时轨迹预测和智能决策来优化交通流、减少拥堵并提高道路安全。因此,本文设计并实现了面向车联网实体相关性分析与多智能体轨迹预测和关键车辆调度优化的应用。本文针对自动驾驶车辆的相关性分析、轨迹预测以及关键车辆的调度决策三个技术进行改革创新。首先分析车辆间的相关性,然后基于车辆的相关性实现了多智能体轨迹预测和关键车辆的调度决策两个应用。通过对车辆进行相关性分析并应用到多智能体轨迹预测与关键车辆调度优化提升交通系统的效率和安全性。具体而言,本文的研究内容可划分为以下三个主要方面。(1)基于对比学习的车辆相关性分析算法基于对比学习研究车联网内车辆节点间的相关性,探讨基于时间和空间的车辆间重叠视野和车辆节点嵌入对相关性的影响,提出了一种基于对比的无监督图表示学习的车辆相关性模型(VC-UGRL),该模型专门用于研究车辆之间的相关性。然后联合轨迹预测下游任务提出了基于对比的无监督图表示学习的车辆相关性的轨迹预测框架,该框架结合图表示学习增强特征提取,能够有效提高下游轨迹预测任务的准确性。(2)基于多智能体交互感知的轨迹预测算法为了预测相关车辆的轨迹,首先提出了一种以场景为中心的基于Transformer的多智能体轨迹预测框架(ABL-Transformer)。该框架通过使用基于注意力机制的Bi-LSTM融合了从过去到未来的持续状态特征并对不同信息进行重要度划分。用以得到更加丰富的信息、精确的环境信息使得车辆可以预见到可能发生的阻碍或者方向的变化,以便于提高轨迹预测精确性。为了进一步降低模型计算开销,根据消融实验的分析结果提出了基于FNet的实时自动驾驶车辆轨迹预测框架(ABL-FNet)。通过采用标准,无参数化傅里叶变换代替Transformer编码器自注意子层的方法,使算法更加轻量,缩短了预测时间,提高了预测效率,同时确保了预测准确率。(3)基于最大权重策略的关键车辆选择优化算法本章基于车辆间的相关性,考虑AoI的演进模型,提出了一种基于最大权重策略的关键车辆选择优化算法(MWP),旨在提高车辆间信息传输的实时性、可靠性和稳定性。通过理论分析和仿真测试,验证了该算法在减少系统平均AoI方面的优异性能。更多还原

【Abstract】 With the progress of artificial intelligence and 5G technology,and the continuous growth of the scale of intelligent networked vehicles,Intelligent Transportation System(ITS)has attracted extensive attention from academia and industry.ITS is an advanced traffic management system that comprehensively applies information,communication and control.Its goal is to improve the efficiency and safety of the traffic system,optimize traffic flow,reduce congestion and improve road safety through real-time trajectory prediction and intelligent decision-making.Therefore,this thesis designs and implements the application of entity correlation analysis,multi-agent trajectory prediction and key vehicle scheduling optimization for vehicle networking.In this thesis,the correlation analysis,trajectory prediction and scheduling decision of key vehicles of autonomous vehicles are reformed and innovated.Firstly,the correlation between vehicles is analyzed,and then two applications,multi-agent trajectory prediction and key vehicle scheduling decision,are realized based on the correlation between vehicles.The correlation analysis of vehicles is applied to multi-agent trajectory prediction and key vehicle scheduling optimization to improve the efficiency and safety of traffic system.Specifically,the research content of this paper can be divided into the following three main aspects.(1)Vehicle correlation analysis algorithm based on contrastive learning.Based on contrastive learning,this thesis studies the correlation between vehicle nodes in the Internet of Vehicles,and discusses the influence of overlapping vision and vehicle node embedding between vehicles based on time and space on the correlation.The vehicle correlation model based on comparison unsupervised graph representation learning(VC-UGRL)is proposed,which is specially used to study the correlation between vehicles.Then,combined with the downstream task of trajectory prediction,a trajectory prediction framework based on vehicle correlation of contrast unsupervised graph representation learning is proposed,which can effectively improve the accuracy of the downstream trajectory prediction task by combining graph representation learning to enhance feature extraction.(2)Trajectory prediction algorithm based on multi-agent interactive perception.In order to predict the trajectory of related vehicles,the scene-centered multi-agent trajectory prediction framework based on Transformer(ABL-Transformer)is proposed.This framework integrates the persistent state characteristics from the past to the future and divides the importance of different information by using Bi-LSTM based on attention mechanism.In order to obtain more abundant information and accurate environmental information,vehicles can foresee possible obstacles or direction changes,so as to improve the accuracy of trajectory prediction.In order to further reduce the computational cost of the model,the real-time trajectory prediction framework for autonomous vehicles based on FNet(ABL-FNet)is proposed according to the analysis results of ablation experiments.By using standard and nonparametric Fourier transform instead of the self-attention sublayer of Transformer encoder,the algorithm is lighter,the prediction time is shortened,the prediction efficiency is improved,and the prediction accuracy is ensured.(3)Optimization algorithm of key vehicle selection based on maxi-weight strategy.Based on the correlation between vehicles and the evolution model of AoI,this chapter puts forward the key vehicle selection optimization algorithm(MWP)based on the maximum weight strategy,aiming at improving the real-time,reliability and stability of information transmission between vehicles.Through theoretical analysis and simulation test,the excellent performance of the algorithm in reducing the average AoI of the system is verified.更多还原