Patch-Based Tracking and Detecting for Visual Tracking

摘要

As one of the most traditional tracking methods, particle filter has been improved in many previous tracking methods due to its non-Gaussian and non-linear distribution. Meanwhile, pure tracking methods cannot achieve good performance in complex tracking scenarios where there enormous deformation and occlusion occur. We present a combination of patch-based tracking and detecting methodology for visual tracking. In our tracking stage, a hierarchical patch-based histogram is used to describe the observation model, computed by an improved L_1 bin-ratio dissimilarity( L_1 -BRD) distance. While in the detecting stage, a patch-based binary feature is obtained through center-symmetric local binary pattern (CS-LBP) and then used to train a randomize fern forest. We combine the two parts collaboratively and experiments demonstrate that the proposed tracking framework outperforms the state-of-the-art methods in challenging scenarios.