Evaluation of Multi-part Models for Mean-Shift Tracking

Abstract

Mean-shift tracking is a data-driven technique for track-

Automatic Detection of Logos in Video and Their Removal Using Inpainting

“…δ is the Kronecker delta function, and b(x i ) is a histogram bin corresponding to pixel value at location . We improve the mean shift tracking method by combining background exclusion [15], [16] and model update [17]. Especially, the background exclusion is also used to extract the logo from the kernel.…”

“…We remove the extracted logo by an inpainting algorithm that is explained in the next step. Second, we subtract the background weight from the foreground weight [15], [16]. Zhao and Nevatia [16] w respectively.…”

A Novel Approach Coloured Object Tracker with Adaptive Model and Bandwidth Using Mean Shift Algorithm

“…In the mean shift tracking algorithm, the colour (Probably Density Function) PDF of target location which is shown by q in colour space and weighted according to an isotropic kernel is target feature [25], pixels closer to the centre of the kernel are assigned greater weight than those near the boundary and the tracker compares two histograms using a metric based on the Bhattacharyya coefficient [52].…”

Track validation using gradient-based normalised cross-correlation

“…Ground truth data was available for the CAVIAR sequences, but it was necessary to create such data for the PETS videos using the ViPER toolkit [6,11]. At each frame, we determine the target's scale by specifying the horizon line in the scene and exploiting the effects of perspective [3][2, pp. 57-60].…”

“…Our gradient-based NCC tracker is, on average, 4.8 times faster than the bruteforce version, while being slightly more robust. 3 (The brute-force trackers must evaluate the similarity measure on a grid of 11 × 11 image locations.) The speed difference is more pronounced for the (lower-resolution) CAVIAR videos, where the gradient-based tracker generally requires fewer iterations to reach a local maximum.…”