Visual tracking in the real world is challenging with unavoidable background interference, target orientation variations and scale changes. Spatial information needs to be exploited to increase robustness; however, current methods such as "Spatiogram" suffer from the large complexity of spatial covariance calculation. Recently, joint distribution representation has been used to estimate target orientation and scale, but this representation is at the expense of losing position localization information. A new framework is proposed for target model representation by employing multiple kernel centers (MKC) within the kernel window. By employing MKC, spatial information is implicitly embedded. Steepest gradient ascent is used to track the target position, orientation and scale simultaneously. Using an adaptive stepsize in the gradient ascent iteration, the proposed method inherits the desirable properties of the mean shift approach and shows a fast convergence rate. The experimental results in several challenging scenarios demonstrate its robustness and superiority to previous technique.
展开▼
