Optimization Algorithm, Kernel-based Video Object Tracking

Visual object tracking has been a hot topic among more and more researchers in recent years. Compared with those tracking methods based on sensors or radars, visual object tracking requires simpler devices, but provides more directly-perceived visual information. As a result, it is widely used in the fields that are closely related to human vision, such as perceptual user interface, video surveillance, traffic detection and vehicle navigation.Among multiple visual tracking algorithms, kernel-based object tracking has attracted a lot of interest because of its lower computational complexity and better performance. And its major problem, the singularity problem has also been paid great attention to.However, another problem called "multiple-extremum issue",which is more general, has not ever been mentioned or noticed. Besides, what kind of kernel can make the tracking process more robust is also worth thinking.Our paper studies the problems above based on the SSD method. In order to solve the multiple-extremum issue, we present a novel kernel-based approach called section-based tracking(SBT) that utilizes the section information provided by the division of the object's weight image.We also illustrate the advantages of SBT based on the definition of entropy. SBT serves to increase the amount of model information, eliminate fake extremal points and make the tracking process more reliable.It can also be naturally extended to deal with scale change and rotation. To further improve the tracking performance, we propose qualitative kernel design rules for SSD tracker and create a new kernel called QPeak. Experiments show that both the SBT method and the new kernel conduce to better performance in tracking process.