基于逐级优化策略的特征退化场景下自动驾驶车辆自主定位方法

Autonomous Localization Method for Autonomous Driving Vehicles in Feature Degradation Scenarios Based on Hierarchical Optimization Strategy

多传感器融合定位是解决卫星遮挡、通信中断等组合导航失效场景下车辆自主定位的有效手段。然而,在环境信息稀疏及相似等场景下,特征退化严重,多传感器融合定位面临巨大挑战。基于此,设计了一种基于逐级优化策略的多传感器融合定位算法,以提高特征退化场景下车辆自主定位精度。首先,为解决实时动态测量技术(Real Time Kinematic,RTK)信号丢失情况下全球导航卫星系统(Global Navigation Satellite System,GNSS)定位精度过低导致初始定位失败的问题,设计融合运动模型和GNSS位置的初始定位算法,提高初始定位精度,实现一级位姿优化;其次,针对特征退化场景中特征点稀疏问题,设计基于平面配准和点云配准融合的位姿优化算法,通过提高稀疏特征点权重以提高定位精度;最后,通过采用误差状态卡尔曼滤波的方式融合惯性测量单元(Inertial Measurement Unit,IMU)数据,输出最终的高精度定位信息。为验证算法有效性,在典型特征退化场景下进行试验测试,试验选取了内蒙古自治区鄂尔多斯市多个矿区场景,分析GNSS定位、LIO-SAM里程计定位、Fast-LIO里程计定位、CT-ICP里程计定位和本文算法定位结果,并与RTK定位结果进行对比。试验结果表明:所提出的改进算法均能对最终结果产生积极影响,算法能够实现较鲁棒的定位。在典型特征退化场景下平均定位误差0.31 m,平均翻滚角误差0.21°,平均俯仰角误差0.52°,平均航向角误差2.93°,相比于其他定位结果,定位误差由米级降低到分米级别,具备明显优势。

Multisensor fusion localization is an effective approach for autonomous vehicle positioning in scenarios where combined navigation fails because of factors,such as satellite obstruction and communication interruptions.However,in cases characterized by sparse environmental information and similarities,severe feature degradation poses significant challenges for multisensor fusion localization.This study proposes a multisensor fusion localization algorithm based on a hierarchical optimization strategy to improve autonomous-vehicle-positioning accuracy in feature degradation scenarios.First,an initial positioning algorithm integrating the motion model and global navigation satellite system(GNSS)position was designed to improve the initial positioning accuracy and achieve first-level pose optimization.Furthermore,a pose optimization algorithm based on plane registration and point cloud registration fusion was designed to address the issue of sparse feature points in feature-degraded scenarios.By increasing the weight of sparse feature points,the algorithm aimed to improve the positioning accuracy.Finally,inertial measurement unit data were fused using an error-state Kalman filter,enabling the output of high-precision positioning information.In this study,experimental tests were conducted in typical feature-degrading scenarios to validate the effectiveness of the algorithm.Multiple mining area scenes from Ordos City,Inner Mongolia Autonomous Region,were selected for the experiments.The GNSS positioning results,LIO-SAM odometry positioning,Fast-LIO odometric positioning,CT-ICP odometric positioning,and positioning results of the algorithm were analyzed and compared with the real-time kinematic(RTK)positioning results.The experimental results indicate that the proposed improvement algorithms positively influence the final results,and that the algorithm can achieve robust localization.In typical feature degradation scenarios,an average positioning error of 0.31 m,average roll angle error of 0.21°,average pitch angle error of 0.52°,and average heading angle error of 2.93°are obtained.Compared with other results,they reduce the positioning error from the meter to decimeter level,demonstrating a clear advantage.