Pairwise Spatiotemporal Partial Trajectory Matching for Co-movement Analysis

TL;DR

This work introduces a novel framework for pairwise spatiotemporal partial trajectory matching that converts city-scale GPS data into time-windowed trajectory images and uses a Siamese network to detect co-movement. By imposing overlap checks and operating on layered temporal views, the method provides interpretable, partial trajectory matching and enables analysis of co-walking routines. Evaluated on a 5-week, 126-participant co-walking study, the approach achieves up to F1 ≈ 0.73 and demonstrates superior performance to traditional time-series baselines at the cost of higher computational load. The technique offers practical value for social-behavior research, urban planning, and health monitoring by revealing when and how often paired individuals share movement patterns, while also highlighting privacy considerations and avenues for efficiency improvements.

Abstract

Spatiotemporal pairwise movement analysis involves identifying shared geographic-based behaviors between individuals within specific time frames. Traditionally, this task relies on sequence modeling and behavior analysis techniques applied to tabular or video-based data, but these methods often lack interpretability and struggle to capture partial matching. In this paper, we propose a novel method for pairwise spatiotemporal partial trajectory matching that transforms tabular spatiotemporal data into interpretable trajectory images based on specified time windows, allowing for partial trajectory analysis. This approach includes localization of trajectories, checking for spatial overlap, and pairwise matching using a Siamese Neural Network. We evaluate our method on a co-walking classification task, demonstrating its effectiveness in a novel co-behavior identification application. Our model surpasses established methods, achieving an F1-score up to 0.73. Additionally, we explore the method's utility for pair routine pattern analysis in real-world scenarios, providing insights into the frequency, timing, and duration of shared behaviors. This approach offers a powerful, interpretable framework for spatiotemporal behavior analysis, with potential applications in social behavior research, urban planning, and healthcare.

Figures (5)

• Figure 1: Proposed method pipeline, where $l$ refers to the number of layers. The steps are labeled (a)-(h). The Siamese Network in (h) is the SigNet model, adapted from signet.
• Figure 2: Detailed view comparing a pair that co-walked on a day (row a) and a pair that did not co-walk on a day (row b). The day is split into one layer images as well as five layer images, where layers 4 and 5 are shown as examples. Both individuals' locations are scaled together within the same latitude and longitude values to fit on the canvas per layer. With five layers, each layer allows for a more detailed view of the locations for each time segment, essentially 'zooming in' to each location visited (e.g., the latest time interval of 19:11-24:00 visualized in red is zoomed in for the five layer scenario as compared to the red trajectory in the one layer scenario).
• Figure 3: Positive class results (F1-score, precision, and recall) from co-walking classification. Each section of the radar charts aligns with each step of the Ablation Study, with $PT$ indicating the Siamese model is pre-trained on signature data, while $FT$ indicates we trained the model on our image trajectory data.
• Figure 4: (a) Routine behavior patterns for two individuals in a pair across all days' of data (Person 1: 23 days, Person 2: 26 days). Utilizing clustering, routine patterns are identified and color-coded by time of day. (b) Five-layer images are generated for each individual from all days of data. Each layer represents a different time window of the day, progressing along the custom gradient. (c) When these layers are passed through a Siamese Network, a distance score is obtained for each layer, indicating the similarity between layers and likely co-walking intervals based on routine behavior. In this case, the pair most likely co-walked during the time interval of Layer 4 since it has the smallest distance, followed by Layer 5. Layer 2 images lacked sufficient data for a reliable distance measure.
• Figure 5: Updated proposed method pipeline, where $l$ refers to the number of layers. The steps are labeled (a)-(f). The Siamese Network in (f) is the SigNet model, adapted from signet. This pipeline reflects the optimized configuration of our initial method based on empirical results.