A comparative study of new HOS-based estimators for moving objects in noisy video sequence

The need for motion estimation (ME) arises quite often in many areas such as computer vision, target tracking, medical imaging, robotic vision. A five new estimators for frame-to-frame image ME are described in this paper. The new ME estimators exploit the higher-order statistics (HOS) characteristics of the received images, and various frequency weighting functions are used to prefilter the received images before calculating the generalized cross-cumulant function and, therefore, suppress the Gaussian noise effect. The estimators of interest are the HOS-ROTH impulse response, the HOS-phase transform, the HOS-smoothed coherence transform, the HOS-maximum likelihood and the HOS-Wiener estimators. Since the performances of the HOS-based estimators are considerably degraded by the signal-to-noise ratio level, this factor has been taken as a prime factor in benchmarking the different estimators. For robust ME it has been found that the HOS-Wiener estimator is particularly suited to this purpose. The accuracy of the estimators is also discussed.