AMGB: Trajectory prediction using attention-based mechanism GCN-BiLSTM in IOV

Accurate and reliable prediction of vehicle trajectories is closely related to the path planning of intelligent vehicles and contributes to intelligent transportation safety, especially in dynamic and uncertain scenarios. However, most existing methods have difficulty in accurately capturing vehicle interactions and the dependencies between vehicle multimodal features in dynamic and uncertain driving environments. Thus, we propose a new Attention-based Mechanism GCN-BiLSTM trajectory prediction model (AMGB) which tackles trajectory prediction in dynamic environments from a new perspective of predicting vehicle motion direction and motion distance. Firstly, the Attention-based Time-Frequency domain Graph Convolutional Network (AT-GCN) module learns the dependencies between vehicle multimodal features and extracts coarse-grained features containing directions information of future trajectories. Then the Multi-structure based Bidirectional Long- Short Term Memory network (M-BiLSTM) module can acquire fine-grained features containing future trajectory distance from vehicle interaction information by the memory storage function of BiLSTM. Finally, we apply the attention mechanism to fuse the coarse and fine-grained features to establish the mapping relationship between vehicle multimodal features and interaction behaviors, and future trajectories. The proposed AMGB model is evaluated on the NGSIM dataset, the results confirm that our model outperforms other state-of-the-art models in both short-long term trajectory prediction.