Multiple kernels have several advantages over single kernel. For, example, multiple kernels can alleviate the singularity and improve the kernel's observability to the motions. Multiple kernels are better at handling tracking an object with complex structure, while a holistic representation based on a single kernel is cumbersome. In such a case, distributing the tracking task into several correlated sub-tasks would be viable. Another benefit is the save of the computation since each sub-task only needs a relatively small kernel.
We think good strategies to place multiple kernels are I)each kernel has a reliable tracking performance, i.e., at a good location, and based on which, II) the structure of the multiple kernels should remain stable through the sequence and be simple.
--- from paper : Efficient Optimal Kernel Placement for Reliable Viusal Tracking
As discussed in the previous section, a single kernel, no matter what its structure, is ultimately limited by two factors: 1) dimensionality of the histogram (which in turn may be a function of available image structure), and 2) the interaction between its derivative structure and the spatial structure of the image as it is exposed by the histogram. Thus, the obvious direction to pursue is to somehow increase the dimensionality of the measurement space, and to simultaneously architect the derivative structure of the kernel to be sensitive to desired directions of motion.
Of course, multiple kernels is not a panacea for improving tracking quality. For example, applying the same kernel at the same location does not improve the rank structure of the system. Similarly, kernels placed or oriented appropriately may not yield independent information. Thus, care and analysis of kernel properties is essential in constructing multi-kernel trackers.
--- from paper : Multiple Kernel Tracking with SSD