This study focuses on the critical aspect of robust state estimation for the safe navigation of an Autonomous Vehicle (AV). Existing literature primarily employs two prevalent techniques for state estimation, namely filtering-based and graph-based approaches. Factor Graph (FG) is a graph-based approach, constructed using Values and Factors for Maximum Aposteriori (MAP) estimation, that offers a modular architecture that facilitates the integration of inputs from diverse sensors. However, most FG-based architectures in current use require explicit knowledge of sensor parameters and are designed for single setups. To address these limitations, this research introduces a novel plug-and-play FG-based state estimator capable of operating without predefined sensor parameters. This estimator is suitable for deployment in multiple sensor setups, offering convenience and providing comprehensive state estimation at a high frequency, including mean and covariances. The proposed algorithm undergoes rigorous validation using various sensor setups on two different vehicles: a quadricycle and a shuttle bus. The algorithm provides accurate and robust state estimation across diverse scenarios, even when faced with degraded Global Navigation Satellite System (GNSS) measurements or complete outages. These findings highlight the efficacy and reliability of the algorithm in real-world AV applications.

翻译：本研究聚焦于自动驾驶汽车安全导航中鲁棒状态估计的关键问题。现有文献主要采用两种主流的状态估计技术：基于滤波的方法和基于图的方法。因子图作为一种基于图的方法，通过构建值节点和因子节点实现最大后验估计，其模块化架构便于集成来自多类传感器的输入。然而当前大多数基于因子图的架构需要预先知晓传感器参数，且仅适用于单一传感器配置。针对这些局限性，本研究提出了一种新颖的即插即用型因子图状态估计器，无需预设传感器参数即可运行。该估计器适用于多传感器配置场景，能以高频提供包含均值和协方差的完整状态估计，且操作便捷。所提算法在四轮车和摆渡车两种车辆平台上，通过多种传感器配置进行了严格验证。实验结果表明，即使面对全球导航卫星系统测量值降级或完全中断的情况，该算法仍能在多样化场景中提供精确鲁棒的状态估计。这些发现凸显了该算法在真实自动驾驶应用中的有效性与可靠性。