车辆和道路的状态是智能汽车安全系统和高阶驾驶辅助系统的重要输入。随着线控底盘和域控制架构的发展，开发统一的全局动力学状态观测方法成为新的趋势。然而，传统分散式底盘电子电气架构由于可用传感器少、算力分散等问题难以满足新型算法开发需求。为此，本研究在底盘动力学域控制架构下，结合底盘域和智驾域传感器信息，开展车路系统关键状态估计研究。首先，设计了融合底盘域传感器和全球定位系统（GPS）的车速和道路坡度估计方法。对纵向车速设计了组合横摆补偿、大小轮速法和双凸点动态斜率法的虚拟传感器，对侧向车速设计了基于长短期记忆神经网络的虚拟传感器。为有效应对大滑移工况，本研究采用逻辑规则和长短期记忆神经网络对车速虚拟传感器的测量噪声方差进行估计，提出了基于长短期记忆神经网络的自适应卡尔曼滤波算法。此外还基于多传感器最优融合算法，实现融合GPS信息的高精度车速估计。为补偿加速度等信号，本研究基于车辆动力学、运动学和GPS信息构建了道路坡度融合观测算法，并基于噪声建模实现自适应权重调整。其次，设计了基于视觉的路面分类算法。本研究构建了包含8种路面类和干扰事物类的路面图像数据集，并设计图像增广方法以匹配更多行驶工况。对车载摄像头采集的图像进行区块划分，使用基于GhostNetV2轻量化卷积神经网络和对比学习框架构建的神经网络进行实时路面分类。基于证据推理算法对区块分类结果进行再融合，以提升分类精度和鲁棒性，得到左右侧路面分类结果。然后，构建了融合车辆动力学和视觉信息的路面附着系数估计策略。基于车辆动力学模型和滑模控制理论设计纵向、侧向轮胎力估计算法，基于载荷转移计算垂向轮胎力。改进Dugoff轮胎模型，并采用主从架构的自适应无迹卡尔曼滤波算法构建基于车辆动力学的路面附着系数估计方案。对路面分类结果进行区间映射和时间同步，获得路面附着系数估计范围。设计可靠性因子以评判车辆动力学和视觉方案的可信度，并基于此构建路面附着系数融合估计策略，设计了基于视觉信息反馈的状态更新方程，提升算法收敛速度，并减轻对动力学激励的依赖。最后，基于搭载底盘域控制器、GPS和摄像头的实车实验平台，在干沥青、冰面、雪面、水泥路面以及混合路面上开展实验，验证了本文所提出的车路系统状态估计算法的有效性及优势。
Vehicle and road states are important inputs to intelligent vehicle safety systems andadvanced driving assistance systems. With the development of X-by-wire chassis anddomain control architectures, there is a new trend to develop unified methods for globaldynamics state observation. However, traditional decentralized chassis architecture is difficult to meet the requirements of novel algorithm development due to the problems offew available sensors and dispersed arithmetic power. Hence, this study carries out thekey state observation research of the vehicle-road system under the chassis dynamics domain control architecture by combining chassis-domain and smart-driving-domain sensorinformation.Firstly, a vehicle velocity and road slope estimation method fusing chassis domainsensors and global positioning system (GPS) is designed. Virtual sensors combining thetransverse yaw compensation, the wheel speed processing method, and the dynamic slopemethod are designed for longitudinal vehicle velocity, and a virtual sensor based on thelong short-term memory neural network is designed for lateral vehicle velocity. In order toeffectively cope with the large slip condition, this study also estimates the measurementnoise variance of the virtual sensors by using logic rules and long short-term memoryneural networks, and proposes an adaptive Kalman filtering algorithm based on long shortterm memory neural networks. Highly accurate vehicle velocity estimation by fusingGPS information is also realized based on the multi-sensor optimal fusion algorithm. Tocompensate for signals such as acceleration, this study constructs a road slope fusionobservation algorithm based on vehicle dynamics, kinematics and GPS information, andrealizes adaptive weight adjustment based on noise estimation.Secondly, a vision-based road classification algorithm is designed. The road imagedatasets of 8 road surface classes and disturbance classes are constructed, and the imageaugmentation methods are designed to match more driving conditions. The images captured by the on-board camera are segmented into blocks, and a neural network constructedbased on GhostNetV2 convolutional neural network and contrastive learning frameworkis used for real-time road classification. The block classification results are re-fused basedon the evidence reason algorithm to improve the classification accuracy and robustness,and the left and right side road classification results are obtained.Thirdly, a tire road friction coefficient estimation strategy fusing vehicle dynamicsand visual information is constructed. Longitudinal and lateral tire force estimation algorithms are designed based on the vehicle dynamics model and sliding mode theory, andvertical tire force is calculated based on weight transfer. The Dugoff tire model is improved. The vehicle dynamics-based road adhesion coefficient estimation scheme usingthe master-slave architecture adaptive unscented Kalman filter algorithm is constructed.Interval mapping and time synchronization on the road classification results is performedto obtain the range of road friction coefficient estimation. Reliability factors are designedto judge the trustworthiness of the vehicle dynamics and vision based methods. Then,the fusion estimation strategy for the road friction coefficient is constructed, and a stateupdate equation based on the visual information feedback is designed to improve the convergence speed of the algorithm and reduce the dependence on the dynamics excitation.Finally, based on the real-vehicle experimental platform equipped with chassis domain controller, GPS and camera, the experiments are carried out on dry asphalt, ice,snow, concrete and mixed road surfaces to validate the effectiveness and advantages ofthe state observation algorithm of the vehicle-road system proposed in this study.
