Abstract:
This research examined an algorithm for filtering unusual GPS data from multiple GPS-equipped
cars. The GPS data was periodically transmitted to a base station via a radio link cvery 3 minutes
and 10 seconds by 3000-4000 taxis around Bangkok and suburb. From manually analysis of the
data, there arc four main problems that affect the accuracy of traffic estimation system: (1) taxi
driver drive round one place for a long time, (2) taxi waits for passengers, (3) the direction of car
does not flow along the street, (4) the measurement error of the GPS receivers. If these problems are
filtered out before being sent to the traffic estimation system, the system will be more accurate. To
validate the performance of the filtered GPS data, the coefficient of variation (C.V.) was applied to
measure the variation of two groups of data: non-filtered data and filtered data. Low value of C.V.
indicates consistent data, thus more accurate. The experiment indicated that the algorithm can
reduce the C.V. of the data by an average of 11.98 percent depending on the locations. Then, a
neural network model was implemented to model the tilter process. The results showed 83.19
percent accuracy compared to the rule-based method.