Efficient Multiscale Hybrid Cross-learning Deep Neural Networks with Spatiotemporal Fusion for Vehicle Behavior Prediction

Recently, deep neural network-based methods have been proposed for vehicle behavior prediction (VBP). However, existing methods lack an enhanced discriminative mechanism and robust learning ability to effectively extract feature information from complex data. This limitation results in information loss, low accuracy, and reduced training efficiency of the model. To address these challenges, a deep convolutional network based on efficient multiscale hybrid cross learning (EMHC) is introduced. This method involves two modules: one module aggregates two independent spatial perception feature maps along different directions and scales, while the other module divides different numbers of channels in parallel to achieve cross-channel feature information fusion. Additionally, local cross-learning mechanisms are incorporated to further enhance effective features. Furthermore, a long short-term memory (LSTM) neural network combined with multihead self-attention (MHSA) is developed to establish short and long-range dependencies of temporal information. This integration aims to accelerate convergence speed and address the vanishing gradient problem. Finally, a spatiotemporal fusion model for vehicle behavior prediction is constructed by combining the aforementioned networks. Comparative analysis with state-of-the-art methods is conducted to showcase the superiority of the proposed model.