频域-时空协同框架下的车辆轨迹预测及意图识别模型OA

To mitigate error accumulation from high-frequency noise and low computational efficiency in complex spatial interaction modeling for trajectory prediction,this paper proposes a frequency-spatiotemporal cooperative prediction framework.The method first applies multivariate variational mode decomposition(MVMD)to historical trajectories,sup-pressing noise and extracting frequency-domain features by constraining anomalous perturbations.Then,convolutional neural networks(CNNs)with feature scaling enhance vehicle-environment interaction perception.Finally,it adaptively fuses multi-frequency trajectory components and spatial interaction features.Combined with LSTM networks and a driving intention probability vector,a mixture density network(MDN)outputs the predicted trajectory's probability density distri-bution.Experiments on the NGSIM dataset show the proposed method outperforms baseline models,effectively co-optimizing frequency-domain analysis,temporal modeling,and interaction perception.The maximum root mean squared error(RMSE)over the 1～5 s prediction horizon is only 1.43 m,meeting complex traffic scenario requirements.