Massively Parallel Discovery of Loosely Moving Congestion Patterns from Trajectory Data

The efficient discovery of significant group patterns from large-scale spatiotemporal trajectory data is a primary challenge, particularly in the context of urban traffic management. Existing studies on group pattern discovery mainly focus on the spatial gathering and moving continuity of vehicles or animals; these studies either set too many limitations in the shape of the cluster and time continuity or only focus on the characteristic of the gathering. Meanwhile, little attention has been paid to the equidirectional movement of the aggregated objects and their loose coherence moving. In this study, we propose the concept of loosely moving congestion patterns that represent a group of moving objects together with similar movement tendency and loose coherence moving, which exhibit a potential congestion characteristic. Meanwhile, we also develop an accelerated algorithm called parallel equidirectional cluster-recombinant (PDCLUR) that runs on graphics processing units (GPUs) to detect congestion patterns from large-scale raw taxi-trajectory data. The case study results demonstrate the performance of our approach and its applicability to large trajectory dataset, and we can discover some significant loosely moving congesting patterns and when and where the most congested road segments are observed. The developed algorithm PDCLUR performs satisfactorily, affording an acceleration ratio of over 65 relative to the traditional sequential algorithms.