Map-free vehicle trajectory prediction method based on heterogeneous graphs and dynamic scene constraints

Predicting the motion of surrounding vehicles is critical for autonomous driving systems. However, existing prediction models relying on high-definition (HD) maps experience degraded performance in scenarios where map data is incomplete or unavailable. To address this limitation, a map-free vehicle trajectory prediction framework, termed Dynamic Scene Prediction (DynaScene-Pred), is proposed. This method extracts implicit road topology priors from historical vehicle trajectories using OPTICS clustering and Bidirectional Long Short-Term Memory (Bi-LSTM) networks to generate virtual lane features, compensating for the absence of explicit map data. A heterogeneous graph is constructed using vehicle and virtual lane nodes. Leveraging graph convolutional networks (GCN) and a multi-dimensional attention mechanism (incorporating spatial, lane, and temporal attention), the model captures complex vehicle-vehicle social interactions and vehicle-lane spatial constraints. A Conditional Variational Autoencoder (CVAE) decoding module is further employed to generate diverse and physically plausible multi-modal trajectories by aligning latent distributions with the inferred scene constraints. Experimental results on the Argoverse benchmark demonstrate that DynaScene-Pred yields competitive performance among map-free baselines. On the validation set, compared to the TR-Pred model, the proposed method achieves relative improvements of 5.7% and 1.8% in [Formula: see text] and [Formula: see text], respectively.