TY - GEN
T1 - Corrections of sensing error in video-based traffic surveillance
AU - Naghiu, Florica
AU - Pescaru, Dan
AU - Magureanu, Gabriela
AU - Jian, Ionel
AU - Doboli, Alex
PY - 2009
Y1 - 2009
N2 - The problem of estimating position of a moving car based on sensor networks was hard investigated over the last period. In this paper we have considered a particle filter design to process the data coming from video sensors and able to predict the next position of a car moving in front of the sensors. The relative error resulting from algorithm will be used to calibrate the surveillance camera, in order to reduce the absolute error. The difference in our approach is that we correct the camera error by trying to predict de driver behavior, based on observing the acceleration of the car. This parameter has considered because it reflects better the behavior of the driver. Thus, in our approach, a car is moving on a road segment: almost constant; smoothly accelerating or decelerating; strongly accelerating; or respective, braking. This behavior is reflected in a probability density matrix of a simplified dynamic Markov Model. Further, the hidden parameter extracted from observation and the probability density values associated with transitions states are passed to particle filter. With the aid of this tool, we predict the next position of the car, and compare it with the observed one. Based on these values the absolute and relative error of the system is computed. As proved by experiments, learning and reducing the relative error will help to reduce the absolute error in car position estimation.
AB - The problem of estimating position of a moving car based on sensor networks was hard investigated over the last period. In this paper we have considered a particle filter design to process the data coming from video sensors and able to predict the next position of a car moving in front of the sensors. The relative error resulting from algorithm will be used to calibrate the surveillance camera, in order to reduce the absolute error. The difference in our approach is that we correct the camera error by trying to predict de driver behavior, based on observing the acceleration of the car. This parameter has considered because it reflects better the behavior of the driver. Thus, in our approach, a car is moving on a road segment: almost constant; smoothly accelerating or decelerating; strongly accelerating; or respective, braking. This behavior is reflected in a probability density matrix of a simplified dynamic Markov Model. Further, the hidden parameter extracted from observation and the probability density values associated with transitions states are passed to particle filter. With the aid of this tool, we predict the next position of the car, and compare it with the observed one. Based on these values the absolute and relative error of the system is computed. As proved by experiments, learning and reducing the relative error will help to reduce the absolute error in car position estimation.
UR - https://www.scopus.com/pages/publications/70350234717
U2 - 10.1109/SACI.2009.5136244
DO - 10.1109/SACI.2009.5136244
M3 - Conference contribution
AN - SCOPUS:70350234717
SN - 9781424444786
T3 - Proceedings - 2009 5th International Symposium on Applied Computational Intelligence and Informatics, SACI 2009
SP - 217
EP - 222
BT - Proceedings - 2009 5th International Symposium on Applied Computational Intelligence and Informatics, SACI 2009
T2 - 2009 5th International Symposium on Applied Computational Intelligence and Informatics, SACI 2009
Y2 - 28 May 2009 through 29 May 2009
ER -