The main topic of this paper is target representation and localization in object tracking.

For real-time processing and implementation reasons, histograms instead of probable density functions (pdfs) are often used to characterize both the target model and the target candidates in object tracking. A direct result of quantization and discretization is the nonsmoothness property embedded in the similarity function, which serves as the indicator of the presence of the object to be tracked. Nonsmoothness hinders the application of fast (for example, gradient-based) search procedures that seek a local maximum.

To circumvent this nonsmoothness problem, this paper introduces an interesting isotropic kernel, with a convex and monotonic decreasing kernel profile, to smooth the similarity function.

Based on the sample estimate of the Bhattacharyya coefficient between the two discrete distributions of target model and target candidates, this paper uses a distance between the two discrete distributions that is metric based, and has many desirable properties compared to other (nonmetric based) measures. The object tracking problem then is reduced to a smooth optimization problem, specifically, the search of the basin of attraction of the local maximum.

Experimental results are also included to validate the new approach. This paper is well written. It presents new and interesting results for efficient object tracking.