Mining Trajectory Patterns Using Hidden Markov Models

Abstract

Many studies of spatiotemporal pattern discovery partition data space into disjoint cells for effective processing. However, the discovery accuracy of the space-partitioning schemes highly depends on space granularity. Moreover, it cannot describe data statistics well when data spreads over not only one but many cells. In this study, we introduce a novel approach which takes advantages of the effectiveness of space-partitioning methods but overcomes those problems. Specifically, we uncover frequent regions where an object frequently visits from its trajectories. This process is unaffected by the space-partitioning problems. We then explain the relationships between the frequent regions and the partitioned cells using trajectory pattern models based on hidden Markov process. Under this approach, an object’s movements are still described by the partitioned cells, however, its patterns are explained by the frequent regions which are more precise. Our experiments show the proposed method is more effective and accurate than existing space-partitioning methods.