Polynomial fitting adaptive Kalman filter tracking and choice of correlation coefficient
Kyle T. Ausfeld, Zoran Ninkov, Paul P. K. Lee, J. Daniel Newman, Gregory J. Gosian · Proceedings of SPIE, the International Society for Optical Engineering/Proceedings of SPIE · 2012
Kalman filters have been used as a robust method for object location prediction in various tracking algorithms for nearly a decade. More recently, adaptive and extended Kalman filters have been employed, making predictions even more reliable. The presented addition to this trend is the employment of a polynomial fit to the history of object locations, using the adaptive Kalman filter framework. This allows the linear state model of the adaptive Kalman filter to predict non-linear motion, making tracking more robust. This modified filter will be used in conjunction with the Mean Shift algorithm as the measurement step. Another important consideration when using a Kalman filter in this manner will be which correlation coefficient is used. The Pearson product-moment correlation coefficient is shown to provide more robust tracking when compared to the Bhattacharyya coefficient when objects have either low resolution or are unresolved.