### BibTeX

@MISC{_experimentalresults.,

author = {},

title = {experimental results.},

year = {}

}

### OpenURL

### Abstract

We here focus on the problem of robust estimation of large dominant image motion under local illumination variations. Of course, this task can be implemented by applying the previous GDIM dynamic image model to the classical direct model-based hierarchical estimation technique. However, this approach discloses the critical limit in the initial parameter estimation process, which will be discussed in detail later. Therefore, we suggest a hybrid hierarchical estimation strategy for accomplishing accurate motion estimation fast and effectively. For this, we first devise two simple and differential dynamic image models for accommodating local illumination variation from the GDIM model through some simple mathematical manipulations. Then, based on each of three dynamic image models, three algorithms for dominant motion estimation are implemented, and they are compared through some tests. As a result, we arrive at the presented hybrid hierarchical estimation framework, which can be briefly described as follows: the three dynamic image models including the GDIM model and the two devised models are employed selectively according to the pyramid level. Especially, since the second devised model is superior to the remaining models in performing the initial parameter updating process at the coarse level, employment of the NDIM dynamic image model for motion estimation at the coarse pyramid level is the essence of the proposed strategy. Finally, experimental results on synthetic and real images are provided to prove the validness and effectiveness of the presented algorithm. 2. Two differential dynamic image models 2 Let mi(x) and s ( x) be the local sample mean i and local sample variance of intensity functions, respectively, within a small neighborhood W centered at current image pixel x at time i. With the assumption of both α(x) and β(x) in the GDIM model being constant 2 within W, substituting the definition of mi(x) and s ( x) into equation (1) gives the following two relations: m x + δ x) = α( x) m ( x) + β( x)

### Keyphrases

experimental result gdim model dynamic image model differential dynamic image model local illumination variation real image devised model critical limit previous gdim dynamic image model dominant motion estimation accurate motion estimation intensity function large dominant image motion initial parameter pyramid level current image pixel coarse pyramid level presented hybrid hierarchical estimation framework simple mathematical manipulation hybrid hierarchical estimation strategy local sample mean presented algorithm robust estimation small neighborhood local sample variance coarse level second devised model ndim dynamic image model motion estimation let mi initial parameter estimation process