Discriminative multi-kernel based hand tracking

Fast motion, background clutter and complicated hand modeling contribute to the challenge of the traditional object tracking. This paper proposes an improved kernel based object tracking method based on discriminative subspace selection and multi-kernel calibration mechanism. First, a bag of candidate subspaces are generated by linear image space mapping, the discriminative subspaces are online selected by evaluating hand and background information model via a discriminative function, then an off-line trained mixture of Gaussian skin model is employed to concentrate subspaces of high confidence to formulate the object representation. Second, to adapt to fast motion and scale variation, the kernel location is calibrated via multi-kernel similarity surface estimation; and facilitated iterative target locating. Experimental results show satisfactory precision as well as real time performance.