Conceptual

STATVTPred: Spatio-Temporal Attention for Target Vehicle Trajectory Prediction

A deep model for predicting a single target vehicle's future path from its GPS trajectory in the Internet-of-Vehicles / Intelligent Transportation Systems setting. It maps the trajectory onto a directed graph and extracts spatial structure with a Graph Attention Network, extracts temporal features from the position sequence with a Transformer, and fuses both with the local road-network structure map to filter and emit a smoothed predicted trajectory. On the T-Drive (Beijing) and Chengdu taxi-trajectory datasets it improves Average Match Rate over a plain Transformer and, more substantially, over an LSTM encoder-decoder baseline.