Trajectory Identification Using a Spatio‑Temporal Siamese Network (ST‑SiameseNet)

The trajectory identification problem aims to verify whether a given set of movement traces originates from a specific individual, which is crucial for applications such as taxi driver authentication, insurance risk analysis, and dangerous driving detection.

Existing approaches often rely on additional sensors or cameras, incurring high costs and limiting scalability, and they typically cannot handle unseen drivers. To overcome these challenges, we introduce a novel framework called ST‑SiameseNet that matches driver identities using only GPS trajectory data.

For each driver, features are extracted from their GPS points (latitude, longitude, timestamp) and fed into a Siamese architecture comprising three shared subnetworks: two LSTM modules (one for occupied‑taxi trajectories, one for empty‑taxi trajectories) and a fully‑connected network (FCN) that learns driver‑specific preferences (e.g., frequent locations, typical start/end times, average trip lengths). The six resulting embeddings are combined via a fully‑convolutional network to produce a similarity score, where values close to 0 indicate the same driver.

We evaluated the model on a dataset of 2,197 Shenzhen taxi drivers collected over ten workdays. GPS records were split into occupied and empty trips, and driver preferences were derived. Experiments compared ST‑SiameseNet against SVM, FNN, and a naïve Siamese baseline, showing that our model consistently outperforms alternatives across all metrics. Ablation studies reveal that incorporating driver preferences improves performance, and that using both occupied and empty trajectories yields the highest accuracy.

Further analysis of generalization demonstrated that increasing the number of training days up to five and the number of drivers up to 500 significantly boosts performance, after which gains plateau. The results confirm that ST‑SiameseNet can reliably identify drivers at scale while requiring only inexpensive GPS data.