STORM: Exploiting Spatiotemporal Continuity for Trajectory Similarity Learning in Road Networks.

Area of Science:

• Computer Science
• Artificial Intelligence
• Data Science

Background:

• Trajectory similarity is crucial for transportation and urban planning but challenging due to varying lengths.
• Current methods use deep learning for trajectory embeddings but neglect spatiotemporal continuity.
• Existing embedding techniques are either task-independent or metric-specific, limiting versatility or effectiveness.

Purpose of the Study:

• To develop a novel trajectory embedding method that combines versatility with metric-specific effectiveness.
• To address the limitations of existing approaches by capturing spatiotemporal continuity in road network trajectories.
• To improve the accuracy and efficiency of trajectory similarity analysis.

Main Methods:

• A two-stage embedding process: initial pre-training decoupled from similarity metrics, followed by metric-specific fine-tuning.
• Trajectory modeling incorporating spatiotemporal continuity via road segment embeddings.
• Utilizing a Transformer encoder enhanced with road network-constrained spatiotemporal semantics.

Main Results:

• The proposed approach demonstrates superior performance in approximating multiple trajectory similarity metrics compared to state-of-the-art models.
• The method effectively captures spatiotemporal continuity, leading to more accurate trajectory representations.
• Experimental results validate the effectiveness of the combined pre-training and fine-tuning strategy.

Conclusions:

• The novel embedding approach offers a versatile yet effective solution for trajectory similarity analysis in road networks.
• Capturing spatiotemporal continuity is key to improving trajectory modeling and similarity approximation.
• This work provides a significant advancement for applications relying on accurate trajectory comparisons.