Clustering Indoor Positioning Data Using E-DBSCAN

Abstract

Indoor positioning data reflects human mobility in indoor spaces. Revealing patterns of indoor trajectories may help us understand human indoor mobility. Clustering methods, which are based on the measurement of similarity between trajectories, are important tools for identifying those patterns. However, due to the specific characteristics of indoor trajectory data, it is difficult for clustering methods to measure the similarity between trajectories. These characteristics are manifested in two aspects. The first is that the nodes of trajectories may have clear semantic attributes; for example, in a shopping mall, the node of a trajectory may contain information such as the store type and visit duration time, which may imply a customer’s interest in certain brands. The semantic information can only be obtained when the position precision is sufficiently high so that the relationship between the customer and the store can be determined, which is difficult to realize for outdoor positioning, either using GPS or mobile base station, due to the relatively large positioning error. If the tendencies of customers are to be considered, the similarity of geometrical morphology does not reflect the real similarity between trajectories. The second characteristic is the complex spatial shapes of indoor trajectory caused by indoor environments, which include elements such as closed spaces, multiple obstacles and longitudinal extensions. To deal with these challenges caused by indoor trajectories, in this article we proposed a new method called E-DBSCAN, which extended DBSCAN to trajectory clustering of indoor positioning data. First, the indoor location data were transformed into a sequence of residence points with rich semantic information, such as the type of store customer visited, stay time and spatial location of store. Second, a Weighted Edit Distance algorithm was proposed to measure the similarity of the trajectories. Then, an experiment was conducted to verify the correctness of E-DBSCAN using five days of positioning data in a shopping mall, and five shopping behavior patterns were identified and potential explanations were proposed. In addition, a comparison was conducted among E-DBSCAN, the k-means and DBSCAN algorithms. The experimental results showed that the proposed method can discover customers’ behavioral pattern in indoor environments effectively.