<p>Vehicle trajectory prediction (VTP) is critical for autonomous driving safety. However, most existing methods are confined to spatial and temporal features, limiting their ability to explore a broader feature space. To address this, we introduce frequency domain analysis for VTP and propose a novel multi-domain perspective method. Our framework effectively integrates frequency, spatial, and temporal features within a “Frequency-Spatial-Temporal" fusion architecture to enhance multi-vehicle trajectory prediction performance. Specifically, we reshape vehicle trajectories into the frequency domain using Fast Fourier Transform (FFT) and amplitude analysis to capture potential periodic patterns. A combination of 2D Convolutional Neural Networks (CNN) and Long Short-Term Memory (LSTM) networks then extracts these frequency features. Moreover, we construct a novel spatial topology graph based on relative vehicle states and actual distances for efficient spatial representation, leveraging stacked Graph Convolutional Networks (GCN) to capture interaction relationships. For temporal dimension, we establish time-dependent connections via a Circular Limited Penetrable Visibility Graph (CLPVG) to obtain rich temporal representations, extracted using a dedicated GCN. Finally, a Gated Recurrent Unit (GRU) encoder-decoder couples these multi-domain features to generate joint predicted trajectories. Extensive experiments on the NGSIM and HighD datasets demonstrate the effectiveness of our approach. It reduces the Root Mean Square Error (RMSE) within a 5-second prediction horizon by 32.13% and 15.45%, respectively, compared to the current state-of-the-art method, and significantly outperforms mainstream baseline models.</p>