融合多率时延测量的无人车位姿估计方法研究

The position and attitude information of ego-vehicle is important for driving of driverless vehicles. The precision and reliability of the systems for position and attitude estimation is vital to safe driving. Fusion visual measurements, which are based on the images of camera and high-precision maps, is promising for high-precision position and attitude estimation due to economic viability. However, the spare sampling intervals and large time-delay of visual measurements make the fusion difficult. It requires that not only the effectiveness and complexity of the estimation model have to be well trade-off but also the algorithm of estimation has to be with high real-time performance. On the other hand, in order to guarantee the reliability, it is important to test the estimation systems. However, it is still challenging how to establish a platform that can test the performance of the estimation systems in all situations. This research addresses all of the mentioned problems. Firstly, multiple models is established that is composed by one kinematics model and several dynamics models. Based on interacting multiple model method, the motion of the vehicles under various driving conditions can be described by a simple way. Secondly, a high real-time algorithm is proposed to fuse large time-delay visual measurements. This method fully utilize the characteristic that the sampling intervals of visual measurements are spare. Finally, an integrating platform is developed that covers the soft simulation, semi-physical simulation and the test of real vehicles. The ability to test and verify the system for position and attitude estimation is improved by applying this platform. Through the in-depth research, the high-precision estimation of position and attitude is achieved and the technical method to test the estimation systems is improved.

无人车的位姿信息是无人驾驶系统进行驾驶操作的重要依据。位姿估计系统的精度和可靠性对行车安全至关重要。融合基于摄像头图像和高精度地图的视觉测量是实现高精度位姿估计的经济可行方案。然而，视觉测量的采样间隔稀疏和大时延特性，不但对平衡估计模型的有效性与复杂性提出更高要求，而且对估计算法的实时性造成挑战。此外，对估计系统的验证是保障其可靠性的重要手段，然而如何建立涵盖各种条件的验证平台仍是亟待解决的问题。针对以上问题，本项目研究车辆运动过程建模方法，建立覆盖不同行驶工况的车辆运动学和动力学多模型，并结合交互多模型方法平衡模型的有效性与复杂性；设计融合大时延测量的低计算复杂度无人车位姿估计算法，以保证估计系统的实时性；研制贯穿软仿真、半实物仿真和实车实验过程的无人车位姿估计一体化验证平台，提高对位姿估计系统的测试验证能力。通过本项目的深入研究，实现无人车高精度位姿估计，完善估计系统验证的技术方法。

本项目研究融合多率传感器信息和延迟视觉测量的无人车位姿的关键问题，旨在提高无人车在各种环境下的定位能力。首先，本项目提出了高精度地图增量式融合更新方法，可以在无人车运行期间持续改善地图的精度，为融合视觉测量的无人车位姿估计奠定了坚实的基础；然后，提出了融合车道线和路沿与车辆相对角度和距离相机视觉测量信息的位姿估计方法，并利用延迟状态增广方法解决了多率和时延测量的融合问题；进而，提出了基于视觉SLAM方法将立体相机与其他车载传感器融合定位的方法，并通过环路闭合检测消除累积误差，有效提高位姿估计精度，可适用于无GNSS信号的情况；更进一步，对于光照条件较差的情况，提出了融合路沿、地面反射强度和环境高度特征的激光视觉测量位姿估计方法；对于环境中存在数量较多动态障碍物从而影响特征匹配的情况，继续研究提出了基于激光视觉空城模型和三维结构特征的无人车位姿估计方法；最后，构建了无人车位姿估计平台，有效提高了位姿估计算法开发、仿真和测试效率，降低开发测试成本和风险。.本项目取得了丰富的理论成果，积累了重要的实验数据，培养了多名研究生。基于项目科研成果，项目组已在知名学术刊物和会议上发表论文16篇，申请发明专利4项。