泊车过程中车辆转向状态的分段位姿估计方法

Segmented Pose Estimation Method for Vehicle Steering State during Parking

针对泊车过程中转向系统非线性因素影响车辆位姿估计准确度的问题,提出了一种基于车辆转向状态的分段位姿估计方法,提高自动泊车过程中的车辆位姿估计精度。首先,通过试验对车辆运动学航迹推算法定位的误差来源进行分析,确定了转向过程的非线性因素对泊车过程中车辆位姿计算准确度的影响;其次,设计了基于车辆转向状态的分段位姿估计模型,使用LSTM对转向状态进行实时分类,分别训练混合模型进行位姿估计,并对分类结果与位姿估计结果进行后验修正;再次,构建泊车工况车辆横纵向位移数据集,基于数据集对所建模型进行训练和离线测试;最后,利用Python搭建在线测试平台,对模型进行位姿估计测试,并开展对比试验。试验结果表明:该方法能够将泊车过程中的车辆欧式距离误差控制在10 cm以内,航向角误差控制在1°以内。基于车辆转向状态的分段位姿估计方法研究,可以有效提高泊车过程中的车辆位姿估计精度,并具有较高的实时性和较好的鲁棒性。

In order to solve the problem that nonlinear factors of steering system affected the accuracy of vehicle pose estimation during parking,a segmented pose estimation method based on vehicle steering state was proposed to improve the accuracy of vehicle pose estimation during automatic parking.Firstly,the error sources of the vehicle kinematic trajectory prediction positioning method were analyzed through experiments,and the influence of nonlinear factors in the steering process on the accuracy of vehicle pose calculation during parking was determined.Secondly,the segmented pose estimation model based on vehicle steering state was designed.LSTM was used to classify the steering state in real time,the hybrid model was trained separately for pose estimation,and posterior correction on the classification results and pose estimation results were carried out.Then,the data set of vehicle transverse and longitudinal displacement under parking condition was constructed.The data set was used for training the established model and offline testing.Finally,Python was used to build an online test platform to test the pose estimation of the proposed model,and a comparative experiment was carried out.The test results show that the European distance error of the vehicle can be controlled within 10 cm and the heading angle error can be controlled within 1°during parking through the proposed method.The segmented pose estimation method based on vehicle steering state can effectively improve the accuracy of vehicle pose estimation during parking and has relatively higher real-time performance and better robustness.