Characterization of Moving Point Objects in Geospatial Data

Summary form only given. Geospatial data that exhibit time varying patterns are being captured faster than we are able to process them. We thus need machines to assist us in these tasks. One such problem is the automatic understanding of the behavior of moving objects for finding higher level information such as goals, intention etc. We propose a system that can solve one part of this complex task: automatic classification of movement patterns made by objects. In addition our system makes some simplifying assumptions: a) the object can be approximated as a moving point object (MPO) b) we consider interaction of a single MPO such as a car or mobile human, with static elements such as road networks and buildings e.g. airports, bus stops etc. on a terrain c) interactions between multiple MPOs are not considered. We use supervised machine learning algorithms to train the proposed system in classifying various patterns of spatiotemporal data. Algorithms such as Support Vector Machines and Decision Tree learning are trained with human labeled feature vectors that mathematically summarize how an MPO interacts with a landmark over time. Our feature vector incorporates a variety of geometric and temporal measurements such as the variable distances of the MPO to different points on the landmark, rate of change with time of variables such as distances and angles that are formed by the MPO with respect to the landmark. Simulated data created through graphical user interaction and agent-based modeling techniques are used to simulate MPO patterns over a representation of a real-world road network. The open source agent-based modeling tool Net Logo along with its GIS extension, and also the Agent Analyst module of ArcGIS are used to simulate large data sets. As future extensions, we are working on classification and prediction problems that involve multiple MPOs and landmarks.