Flexible Contact Correlation Learning on Spatio-Temporal Trajectories

Chenhao Wang, Lisi Chen, Shanshan Feng, Shuo Shang

Abstract: The study of contact between moving objects has received increasing attention in recent years. Existing studies have primarily focused on contact search or tracing, aiming to identify all trajectories in contact with a query trajectory. However, these studies only consider spatial contacts at specific timestamps, and highly rely on precise data with consistent sampling rates and aligned timestamps. In light of these limitations, we investigate the problem of trajectory contact correlation learning, which represents a continuous evaluation of the spatio-temporal contact between trajectories. In addition, we introduce a novel approach called the Spatio-Temporal Trajectory Contact Network (ST-TCN) for learning contact scores between trajectory pairs. Specifically, the ST-TCN first computes pointwise contact weights using a proposed contact attention module, followed by the identification of potential contact positions using a soft selection module. The contact scores are then derived from the embeddings of the contact trajectory parts. Experiments on two real-world datasets show that ST-TCN outperforms baseline solutions and exhibits superior efficiency in terms of both running time and GPU memory usage. The source code can be found at https://github.com/chwang0721/ST-TCN.