Analyzing Abnormal Events from Spatio-Temporal Trajectories Dhaval Patel, Chidansh Bhatt, Wynne Hsu, Mong Li Lee, Mohan Kankanhalli School of Computing National University of Singapore - Singapore {dhaval,chidansh,whsu,leeml,mohan}.comp.nus.edu.sg Abstract—Advances in RFID based sensor technologies has been used in applications which requires the tracking of assets, products and individuals. The recording of such movements is captured in a trajectory database and can be analyzed for the monitoring of abnormal events. In this paper, we describe a system called InViTA for analyzing abnormal events from spatio-temporal trajectories captured during an ofſce evacuation after an explosion. InViTA utilizes a trajectory representation scheme and extract the features to derive a set of rules that label each person’s trajectory as belonging to a suspect, witness, or victim, etc. We run the system on the ofſce evacuation data provided in VAST 2008 challenge and obtain comparable results with that obtained from visualization and human analysis. The system includes a user-friendly graphical interface for parameter tuning and intuitive result analysis. Keywords-representation scheme, trajectory classiſcation, abnormal events I. I NTRODUCTION Many government organizations such as embassies, multi- nationals companies, shopping malls, banks and etc use surveillance systems to record movements of employees, visitors, merchandise etc. Video cameras or RFID tags are frequently utilized to record the locations of each employee and visitor at regular time points, that is, these devices capture the spatio-temporal trajectories of each individual. Figure 1 gives an example of such trajectories recorded using RFID tags. These trajectories can be analyzed in two modes. In real time or online processing, the objective is to detect suspicious people based on their movements in order to circumvent any potential security threats and minimize loss or casualty [2]. Off-line processing is typically focused on analyzing the trajectories before and after the occurrence of an abnormal event in order to obtain more insights into causes and characteristics of the event. An abnormal event includes an explosion in a building, open shooting in a campus, hit-and-run accident, etc. Such events will trigger investigations to identify suspects, victims and witnesses. Existing approaches mostly utilize animation based visualization techniques to display all the trajectories and a person is needed to manually select trajectories which seems to be suspicious [3]. However, plotting trajectories in a two-dimensional space is limited since the paths of two persons could be the same but their speeds along the paths could vary greatly. For example, in the case of an explosion, the majority of the people would run for the nearest exit, 10 20 30 40 50 60 70 80 90 10 20 30 40 50 60 Figure 1. RFID Trajectories on Building Floor Map while the terrorist responsible for planting the bomb may walk slowly to a safe exit. Further, the manual determination of suspicious trajectories is not scalable and prone to human errors and subjectivity. In this paper, we describe a system to systematically analyze trajectories and identify suspicious people, witnesses and locate casualties in an ofſce evacuation scenario. Our system overcomes the limitation of two-dimensional trajec- tory plots by transforming two-dimensional trajectories into two one-dimensional time series. This allows us to differ- entiate trajectories with different speeds passing through the same region, as well as characterize the trajectories before and after the occurrence of an abnormal event. This char- acterization reƀects the drastic change in the environment as a result of an abnormal event, thereby allowing us to pinpoint the time and location of the event. Based on the time and location of the abnormal event, we use a set of rules and develop an effective and accurate technique to label the trajectories of suspicious persons, identify witnesses and locate casualties. The contributions of this work are summarized as follows: • Transformation Scheme: We design two measures to transform the two-dimensional spatio-temporal trajecto- ries into two one-dimensional time series. The measures are Window based Average Speed (WAS) and Window based Average traversed Region (WAR). The transfor- mation facilitates the visualization of trajectories of varying speeds and provides for the detection of the occurrence of an abnormal event. • Trajectory Classiſcation: We devise an effective and accurate trajectory classiſcation method based on the 2009 IEEE International Conference on Data Mining Workshops 978-0-7695-3902-7/09 $26.00 © 2009 IEEE DOI 10.1109/ICDMW.2009.45 630 2009 IEEE International Conference on Data Mining Workshops 978-0-7695-3902-7/09 $26.00 © 2009 IEEE DOI 10.1109/ICDMW.2009.45 638 2009 IEEE International Conference on Data Mining Workshops 978-0-7695-3902-7/09 $26.00 © 2009 IEEE DOI 10.1109/ICDMW.2009.45 616 Authorized licensed use limited to: University of North Texas. Downloaded on February 15,2010 at 12:20:04 EST from IEEE Xplore. Restrictions apply.