环境感知结果的准确性和可靠性对于提升自动驾驶的安全和通行效率至关重要。然而，环境感知系统受到内外部多种因素的影响而存在多种不确定性。如何在线量化评估感知系统的不确定性并实现在线处理，兼顾自动驾驶的安全性和效率，成为自动驾驶技术发展的重要需求。本文围绕上述问题展开研究。首先，构建了感知系统不确定性的六层表达模型，对感知不确定性的内涵进行了全面的剖析和表达，包括传感器数据质量、标定参数、感知结果存在性、语义分类、空间信息和运动信息的不确定性，提出了多源信息互验的感知系统不确定性在线量化评估与处理架构，为后续工作奠定理论基础。其次，提出了面向预期功能安全的传感器性能不确定性在线量化评估方法。提出了导致传感器数据质量下降的关键触发因素的在线识别方法，实现了雨天、雪天、雾霾等异常状况严重程度的准确识别，支撑了传感器数据质量的在线量化评估；提出了基于轨迹匹配与特征匹配的传感器标定参数残差在线评估方法，支撑了行驶过程中对标定参数不确定性的在线量化评估。再次，提出了基于“先验-互验”模型的感知结果不确定性在线量化评估方法。提出了利用地图先验信息的静态要素感知结果不确定性在线量化评估方法，提出了基于深度集成的动态要素感知结果不确定性在线量化评估方法，并建立了静态要素和动态要素的不确定性的关联与推理机制，进一步提出了基于多源信息交叉匹配的感知结果不确定性综合评估方法。然后，提出了基于边界式安全行驶空间的不确定性在线处理方法。提出了基于感知结果不确定性的注意力分配机制，降低高不确定性感知结果在多源信息融合过程中的权重；构建统一表达感知结果不确定性的边界式安全行驶空间，实现了感知系统不确定性所带来风险的在线评估。最后，在北京市高级别自动驾驶示范区、清华大学周边道路及首钢冬奥园区进行了实车试验，并结合仿真数据对本文方法进行验证。结果表明，本文方法可以在线量化评估自动驾驶感知系统的各项不确定性，具有比现有评估方法更好的准确性和完整性，所构建的边界式安全行驶空间使车辆在运行中的平均接管距离和运行速度分别提升了 31.69% 和 9.85%。

The accuracy and reliability of environmental sensing system are crucial for enhancing the safety and traffic efficiency of autonomous driving. However, a range of internaland external factors can affect environmental sensing systems, introducing various uncertainties. Quantifying and processing the uncertainties in sensing systems, while ensuringthe safety and efficiency of autonomous driving, is crucial for advancing autonomousdriving technologies. This paper conducts research around the above issues.Firstly, a six-layer expression model of sensory system uncertainty was constructedto provide a comprehensive analysis and expression of the connotation of sensory uncertainty, including the uncertainty of sensor data quality, calibration parameters, existence of sensing results, semantic classification, spatial information, and motion information. An online quantification and processing architecture for sensory system uncertainty,which verifies multi-source information, was proposed to lay the theoretical foundationfor subsequent work.Secondly, an online quantification method for sensor performance uncertainty aimedat expected functional safety was proposed. An online identification method for key triggering factors leading to the degradation of sensor data quality was introduced, achievingaccurate recognition of the severity of abnormal conditions such as rain, snow, and haze,supporting the online quantification assessment of sensor data quality. An online evaluation method for sensor calibration parameter residuals based on trajectory matchingand feature matching was proposed, supporting the online quantification assessment ofcalibration parameter uncertainty during driving.Thirdly, an online quantification method for the uncertainty of sensing results basedon a “prior-verification” model was proposed. A method was introduced that utilizesmap prior information for the online quantification assessment of the uncertainty in staticelement perception results and a deep integration-based method was proposed for theonline quantification of uncertainty in dynamic element sensing results. A mechanismfor the association and reasoning of uncertainties between static and dynamic elementswas established, further proposing a comprehensive evaluation method for sensing resultuncertainty based on multi-source information cross-matching.Then, an online processing method for uncertainty based on boundary-based safe driving space was proposed. An attention distribution mechanism based on the uncertainty of sensing results was introduced to reduce the weight of high uncertainty sensingresults in the process of multi-source information fusion. A unified expression of sensingresult uncertainty in a boundary-based safe driving space was constructed, achieving anonline assessment of risks brought by perception system uncertainty.Finally, real-vehicle experiments were conducted in Beijing High-Level AutomatedDriving Demonstration Zone, around Tsinghua University, and in the Shougang WinterOlympics Park, combined with simulation data to validate the methods of this paper.The results show that the methods proposed in this paper can quantify the uncertainties of the autonomous driving perception system, with better accuracy and completenessthan existing evaluation methods. The constructed boundary-based safe driving space hasincreased the average takeover distance and running speed of vehicles during operationby 31.69% and 9.85%, respectively.