Deep Learning Method Based on Multiscale Enhanced Feature Fusion for Vehicle Behavior Prediction

Vehicle behavior prediction (VBP) is a crucial area of research aimed at enhancing the safety of intelligent connected vehicles in complex traffic scenarios. However, many existing methods rely on 1-D driving data or lack the fusion processing of multiscale input features, leading to reduced model prediction accuracy and unnecessary computational load. In this study, we introduce a deep neural network for multiscale enhanced feature fusion. Initially, we utilize Gramian angular field (GAF) and exponential short-time Fourier transform (ESTFT) to transform 1-D data into a graph containing time-frequency information, thereby improving the interpretability of input features. Subsequently, we develop a multiscale cross-learning attention (MCA) mechanism to facilitate cross-domain and cross-scale feature interactions within the 2-D graph, along with a multiscale global attention (MGA) mechanism to enhance feature learning by considering the spatial relationships of 1-D data. Finally, we utilize multi head attention mechanism to allocate adaptive weights to multiscale enhanced feature information, and incorporate bi-directional long short-term memory (Bi-LSTM) to model time dependencies and generate prediction results. Our proposed approach demonstrates improved performance in VBP compared to existing methods, as validated on both HighD and NGSIM datasets.