Abstract
This study presents a novel social-goal attention networks (SGANet) model that employs a vision-based multi-stacked neural network framework to predict multiple future trajectories for both homogeneous and heterogeneous road users. Unlike existing methods that focus solely on one dataset type and treat social interactions, temporal dynamics, destination point, and uncertainty behaviors independently, SGANet integrates these components into a unified multimodal prediction framework. A graph attention network (GAT) captures socially-aware interaction correlation, a long short-term memory (LSTM) network encodes temporal dependencies, a goal-directed forecaster (GDF) estimates coarse future goals, and a conditional variational autoencoder (CVAE) generates multiple plausible trajectories, with multi-head attention (MHA) and feed-forward networks (FFN) refining the final multimodal trajectory prediction. Evaluations on homogeneous datasets (JAAD and PIE) and the heterogeneous TITAN dataset demonstrate that SGANet consistently outperforms previous benchmarks across varying prediction horizons. Extensive experiments highlight the critical role of socially-aware interaction weighting in capturing road users' influence on ego-vehicle maneuvers while validating the effectiveness of each network component, thereby demonstrating the advantages of multi-stacked neural network integration for trajectory prediction. The dataset is available at https://usa.honda-ri.com/titan.