Efficient multi-target vehicle trajectory prediction based on multi-scale graph convolution

Multi-target vehicle trajectory prediction holds significant importance in the field of autonomous driving. Accurate trajectory prediction can enhance the safety and efficiency of autonomous vehicles, reducing traffic accidents and congestion. However, existing methods often fall short when dealing with complex traffic scenarios. Traditional approaches typically rely on single-scale spatiotemporal feature extraction, which struggles to fully capture the complex dynamics of traffic across different temporal and spatial scales, especially in high-density traffic environments. To address these challenges, this thesis proposes a multi-target vehicle trajectory prediction method based on a Multi-Scale Graph Convolutional Network (MSGCN). This method integrates high-definition semantic maps and employs a spatiotemporal multi-head attention mechanism alongside an adaptive dynamic weighting module to achieve efficient multi-target vehicle trajectory prediction. Specifically, this thesis constructs a dynamic feature repository using vehicle subgraphs and lane subgraphs to stabilize model weight fluctuations, thereby more accurately reflecting actual traffic conditions. Experimental results on the Argoverse dataset demonstrate the effectiveness of our method. Specifically, our approach reduces the average displacement error (mADE) by 7% and enhances the final displacement error (mFDE) by 19% when compared to existing state-of-the-art models. Our code is made available at https://github.com/Garegreen/EfficientMSGCN.