【摘要】 为实现左右舵不同驾驶习性驾驶人在港珠澳大桥时空混行环境下快速、有效识别并预测行驶车辆在应急条件下的运动状态，提出了一种考虑右舵驾驶行为的模型加数据混合运动预测方法。首先，提取港珠澳大桥通行车辆的跟驰与换道原始轨迹数据并分析，挖掘左右舵驾驶行为在直道及变道属性下的长短时特性；其次，结合最大信息系数算法(MIC)对比所提取特征与2类驾驶行为的关联程度，并求解关键区分特性下高斯混合模型(GMM)对于左右舵驾驶行为应急反应的倾向性概率；最后，将2种驾驶行为的车辆运动状态在直道行驶的差异特征作为长短时记忆(LSTM)神经网络的输入，建立数据驱动下的直道横向偏移预测模型，并在具有差异化驾驶行为的车辆直道位姿信息预测基础上，串联建立模型驱动下的变道概率预测模型。对青州航道桥实际车流监测数据的测试结果表明：所提方法可基于行驶车辆的横向偏移和偏航率等特征快速、准确识别左右舵驾驶行为；对于不同特征输入下的直道偏移预测结果，所预测左舵驾驶行为的均方根误差(RMSE)、改进的豪斯多夫距离(MHD)与决定系数(R~2)的最优评估分别为0.578 7、0.468 1与0.870 7,右舵驾驶行为预测结果评估分别为0.711 9、0.588 4与0.864 4;以换道前2 s对左右舵驾驶行为变道意图的预测准确率分别为83.0%和81.3%。该方法后续可以推广到更多左右舵时空混行环境中，利用行驶车辆的横向位姿等长短时特征，精细化预测并诱导存在差异化驾驶行为的紧急避险行为，提高车辆的交互安全性与效率。更多还原

【Abstract】 To improve the safety and efficiency of vehicles, the rapid and effective recognition and prediction of the emergency response motion state of vehicles is an important feature of predictive models. To achieve this goal, a hybrid motion prediction method based on both model and experimental data and considering right-rudder driving behavior is herein proposed. It considers drivers with different driving habits in the spatiotemporal mixed traffic environment of the Hong Kong-Zhuhai-Macao Bridge, as an example. First, the original track data of the car following path and lane changing of vehicles passing through the Hong Kong-Zhuhai-Macao Bridge were extracted and analyzed, and the long-term and short-term characteristics of left-right rudder driving behavior in straight and lane-changing motions were mined. Second, the maximum information coefficient(MIC) algorithm was used to compare a possible correlation between the extracted features and two types of driving behaviors; the tendency probability determined by the gaussian mixture model(GMM) for the emergency response of left-right rudder driving behaviors under the key distinguishing characteristics was calculated. Third, a data prediction model of the straight road lateral offset was established by inputting the difference characteristics of vehicle motion states of the two driving behaviors under analysis on a straight road into the long short time memory(LSTM) neural network. Based on the prediction of the straight road position and attitude information of vehicles, with different driving behaviors, a model of the lane-changing probability was established in series. The test results of the actual traffic flow monitoring data from the Qingzhou Channel Bridge show that the proposed method can quickly and accurately recognize the left-right rudder driving behavior based on the lateral offset and yaw rate of running vehicles. For the straightway offset prediction results under different feature inputs, the optimal evaluation fit of the root mean square error(RMSE), modified Hausdorff distance(MHD), and deterministic correlation coefficient(R~2) of the predicted left-rudder behavior are 0.578 7, 0.468 1 and 0.870 7, respectively. The predicted results for right-rudder behavior are 0.711 9, 0.588 4 and 0.864 4, respectively. The prediction accuracy of lane-changing intention of left-right rudder driving behavior in the 2 s prior to lane change, as set in this paper, is 83.0% and 81.3%, respectively. This method can be further extrapolated to fit other left-right rudder spatiotemporal mixed traffic environments. Through this method, emergency avoidance behavior with differentiated driving behavior can be accurately predicted and deduced to improve the interaction safety and efficiency of vehicles using long-term and short-term characteristics, such as the lateral position and posture of moving vehicles.更多还原