Contents I. Introduction
With the increasing popularity of GPS and RFID devices, the tracking of motor vehicles and moving objects in general has become a reality in many cities. One very practical and important problem in vehicle traffic analysis is outlier detection. A typical definition of an outlier is “an observation (or a set of observations) which appears to be inconsistent with the remainder of that set of data [2].” This rather vague definition can lead to many different outlier detection algorithms. This work will primarily use the application scenario of detecting temporal outliers in vehicle traffic data. More specifically, it seeks to detect outlier behavior in the set of road segments of the traffic data and not individual moving objects. Consider only the solid line in Figure 1, which shows a toy example of load (count) of vehicles on road segment over the course of several days. From this line alone, there does not seem to be any abnormal behavior. But suppose additional information about other road segments in the city are given. In light dashed lines are many other road segments that have similar loads as road segment from July 1 st to July 4th. On July 4th, however, they all show an increase in load while _ remains the same. The question of outliers has now become much trickier. has now just become an outlier. Recall the outlier definition regarding “inconsistency.” This example and the rest of this work use historically similar neighbors as the basis for consistency comparisons. Historical Load on Many Road Segments
