【作者】 刘晶；

【导师】 丁玎； 李福存；

【作者基本信息】 东南大学 ， 软件工程（专业学位）， 2025， 硕士

【摘要】 车辆轨迹预测技术是当前高级驾驶辅助系统发展的关键组成部分,对保障道路交通安全具有重要意义,而驾驶员作为车辆行为的决策主体,其身心状态也对车辆轨迹的演化产生显著影响。随着显示与仿真技术的不断发展,越来越多的研究开始依托驾驶仿真平台开展驾驶行为分析,而虚拟现实技术的发展为驾驶仿真平台的构建提供了更为安全与高效的实现途径。然而,在驾驶员状态检测与车辆轨迹预测领域仍面临两大挑战:(1)如何充分融合用于状态检测的多模态多变量时间序列;(2)如何围绕目标车辆高效建模状态感知与动态交互。针对上述问题,本文基于虚拟现实环境,以实现精确驾驶员身心状态检测以及准确车辆轨迹预测为目标任务开展了研究,具体包含以下四部分:(1)针对数据集构建,开展了基于驾驶仿真平台的数据采集实验,构建了多模态模拟数据集,为模型的训练与验证提供了数据支撑。(2)针对驾驶员身心状态检测,提出了基于深度时空特征融合的驾驶员身心状态检测方法。通过构建多模态时空特征融合网络,深入建模模态内部的上下文信息与模态之间的互补关联。同时,联合驾驶员压力识别与疲劳检测两个任务,实现对驾驶员身心状态较为全面的建模与精准检测。实验结果显示,所提方法在压力识别任务中分别达到81.78%的准确率与82.57%的F1分数,在疲劳检测任务中分别实现89.07%的准确率与88.54%的F1分数,整体性能优于现有主流方法。(3)针对车辆轨迹预测,提出了基于状态感知与交互建模的车辆轨迹预测方法。通过构建基于星型图的状态感知网络,对目标车辆的外部运动状态及其与邻车的动态交互关系进行建模,并引入驾驶员身心状态特征进一步增强状态表征。随后,联合模拟数据与真实数据对网络模型进行端到端优化,有效提升轨迹预测的准确性与鲁棒性。实验结果显示,所提方法在真实场景下1至5秒的平均位置误差分别为0.42、1.04、1.81、2.79和4.03米,整体精度优于现有主流轨迹预测方法。(4)基于上述理论成果,本文设计并实现了基于沉浸式虚拟现实的车辆轨迹预测系统,完成了核心功能模块的开发以及整体系统的部署,并对系统性能与可用性进行了测试,验证了该系统对于检测驾驶员身心状态及预测车辆轨迹的准确性。综上所述,本文围绕核心任务,提出了基于深度时空特征融合的驾驶员身心状态检测方法,以及基于状态感知与交互建模的车辆轨迹预测方法。实验结果表明,所提出的方法在任务精度与泛化能力方面均优于现有主流方法。本文的研究成果有助于推动智能驾驶辅助系统的发展,具有一定的工程应用价值和现实意义。更多还原

【Abstract】 Vehicle trajectory prediction is a critical component in the advanced driver assistance systems(ADAS)and plays a crucial role in enhancing road safety.As the decision-maker in vehicle control,the driver is significantly influenced by their psychophysiological state,which in turn affects trajectory evolution.With advances in visualization technologies,driving simulation platforms are increasingly employed to study driving behaviour.Meanwhile,virtual reality(VR)has provided safer and more efficient approaches for developing simulation platforms.However,challenges remain in two key areas:(1)how to effectively fuse multimodal and multivariate time series for state detection; and(2)how to efficiently model state awareness and dynamic interactions centered around the target vehicle.To address above challenges,this thesis conducts research within a virtual environment,with the dual objectives of accurately detecting drivers’ psychophysiological states and precisely predicting vehicle trajectories.The main work includes four parts:(1)For dataset construction,a data collection experiment was conducted on the driving simulation platform to construct a multimodal simulated dataset,providing data support for subsequent model training and validation.(2)Regarding driver psychophysiological state detection,a Multimodal Spatial-Temporal Feature Fusion Network(MSTF2Net)is proposed.MSTF2 Net effectively captures both intramodal contextual information and inter-modal complementary relationships during the fusion process.Furthermore,it jointly learns stress recognition and fatigue detection to enable comprehensive modeling and accurate assessment of drivers’ psychophysiological states.Experimental results demonstrate that MSTF2 Net achieved the best accuracy and F1-score for both stress recognition(81.78%,82.57%)and fatigue detection(89.07%,88.54%).(3)For vehicle trajectory prediction,this thesis proposes a Star Graph-based State-aware Vehicle Trajectory Prediction(S2GVTP).S2 GVTP models the external motion state of the target vehicle and its dynamic interactions with neighboring vehicles,while also integerating psychophysiological features encoded by MSTF2 Net to enhance state representation.The network is optimized by jointly leveraging simulated and real-world datasets,thereby improving the accuracy and robustness of trajectory prediction.Experimental results show that S2 GVTP achieves root mean square errors of 0.42,1.04,1.81,2.79,and 4.03 meters at 1 to 5 seconds,outperforming existing baseline models.(4)Based on the above theoretical findings,this thesis designs and implements a VR-based vehicle trajectory prediction system.The thesis completes the development of core functional modules,deploys the overall system,and tests the performance and usability of the system,validating its accuracy in detecting drivers’ psychophysiological states and predicting vehicle trajectories.In summary,targeting the research objectives,the thesis proposes MSTF2 Net for driver state detection and S2 GVTP for vehicle trajectory prediction.Experimental results demonstrate that the system has high usability,and the methods outperforms state-of-the-art approaches in both task accuracy and generalization capability.The findings of this thesis contribute to the advancement of ADAS and hold significant potential for practical engineering applications.更多还原