DidaWiki

The widespread use of positioning technologies ranging from GSM and GPS to WiFi devices tend to produce large-scale datasets of trajectories, which represent the movement of travelling entities. Several application domains, such as recreational area management, may benefit from analysing such datasets. However, analysis results only become truly useful and meaningful for the end user when the intrinsically complex nature of the movement data in terms of context is taken into account during the knowledge discovery process. For this reason we propose a pattern interpretation framework that consists of three main steps, namely, pattern discovery, semantic annotation and pattern analysis. The framework supports the understanding of movement patterns that were extracted using some trajectory mining algorithm. In order to demonstrate the feasibility and effectiveness of the framework, we have specifically applied it for understanding moving flock patterns in pedestrian movement. For the pattern discovery step, we have formally defined the concept of moving flock, distinguishing it from stationary flock, and developed a detection algorithm for it. A set of guidelines for setting the parameters of the algorithm is provided and a specific technique is implemented for the radius parameter. As for the semantic annotation step, we have proposed a guideline for selecting appropriate attributes for semantic enrichment of individual entities and of moving flocks. Two levels of annotation, which are at individual and pattern level, were also described. Finally, for the pattern interpretation step, we have combined the results obtained using hierarchichal clustering and decision tree classification in order to analyse the attributes of flocks members and of the flocks, and the flocks themselves. The entire framework was tested on the Dwingelderveld National Park (DNP) dataset and the Delft dataset, both of which are pedestrian datasets based in the Netherlands. The DNP dataset contains records of observations on the movement of visitors in the park while the Delft dataset describe movement of the pedestrians in the city. As a result, some forms of interactions, such as certain groups of visitors following the most popular path in the park, were inferred. Furthermore, some flocks were linked with specific attractions of the park.