Traffic State Estimation under Uncertain Sensor Data
Capturing the traffic system characteristics in real
time is one of the basic requirements of any Intelligent Transportation Systems
(ITS) applications. With the advancement in the ITS sector, several types of
nonintrusive traffic sensors using latest technologies are available for
vehicle sensing. In order to make the task of data collection, processing and
providing real time information to users faster, there is a need for automating
the associated processes. However, these processes are highly challenging under
Indian traffic condition, which is heterogeneous in nature and lacks lane
discipline. The automated sensors, which are working well elsewhere, may not be
able to perform accurately under such conditions leading to the automated data
to be highly erroneous. This in turn would result in reduced performance of the
end application and hence a reduction in accuracy and reliability of the end
application is a major concern. The present study addresses these challenges,
taking the real time estimation of speed and density as sample applications.
Artificial Neural Networks (ANN), which is a data driven approach, and a
dynamical systems based approach using the Kalman
Filter (KF), which is less data demanding in terms of data, were selected as
the techniques for estimation. With ANN, location based speed estimation was
carried out using data from two different sensors representing erroneous and
non-erroneous data. Stream based density and speed estimation were also carried
out while using erroneous data as inputs to a network trained with accurate
data. For the dynamical systems approach based method, density was estimated
using the conservation principle and speed was estimated using a dynamic speed
equation formulated by incorporating appropriate speed-density relationship for
the specific traffic under study. In order to corroborate the developed
estimation schemes, actual density values calculated using manually extracted
flow data and speed that were obtained from travel time data collected by using
Bluetooth sensors were used.
The results indicate that if the offline training of
ANN can be carried out with correct data, ANN can take into account the errors
in inputs. The estimation accuracy of the dynamical systems approach based
method was checked before and after incorporating the sensor error statistics
into the estimation scheme. The results indicated a significant improvement in
the estimation accuracy after incorporating the sensor error statistics. On
overall comparison between two approaches, speed estimation accuracy using ANN
and the dynamical systems approach based method with error statistics were
comparable for both shorter and longer durations. On the other hand, density
estimation accuracy for shorter durations using the dynamical systems approach
based method with error statistics was better. From this study, it was observed
that both ANN and the dynamical systems approach based method can deal with the
erroneous automated sensor data and are able to provide reasonably accurate
results. This shows a promising possibility for user agencies to implement such
applications even if the real time data source is erroneous.