Robust Object Tracking Using Adaptive Multi-Features Fusion Based on Local Kernel Learning

This paper presents a novel multi-features fusion tracking algorithm based on local kernels learning. Histograms of multiple features are extracted based on sub image patches within the target region, and the features fusion weights are calculated respectively for each patch according to the discriminability of features. It means that the same feature employed in different sub image patches gets different weights. In this way, more precise features fusion weights are provided which lead to a more accurate tracking localization. Moreover the spatial information introduced by the sub patches enhances the tracking robustness. A formula for target localization with adaptive multi-features fusion based on local kernels is deduced. Experiments on challenging video sequences demonstrate that the proposed tracking algorithm performs favorably against trackers using usual target representation, without increasing significantly the computational complexity.