Transformer-based Dynamic Interaction Graph Convolution Method for Vehicle Trajectory Prediction

Vehicle trajectory prediction plays a crucial role in autonomous driving, as dynamic interactions between vehicles make trajectory prediction challenging. To more accurately understand the spatial dynamic interactions between vehicles to advance autonomous driving technology, a transformer-based dynamic interaction graph convolution method for vehicle trajectory prediction is proposed. Our model first utilizes an encoder and multi-head attention mechanism to extract comprehensive historical trajectory features of vehicles. Concurrently, graph convolution and observation tensors are employed to capture the dynamic spatial relationships between target and surrounding vehicles. Next, we employ a Transformer model to integrate temporal features and enhanced spatial interaction features, generating future trajectory predictions for the target vehicle through a decoder. Finally, we evaluate our proposed model on the NGSIM datasets, with experimental results demonstrating superior outcome compared to other methods in most scenarios.