Visual tracking using Blind Source Separation for mixed images

Robust tracking using visual cue integration for mobile mixed images

“…To improve tracking accuracy, the [I, R-G, Y-B] color histogram is computed for the intersection of the region indicated by the dynamic layer, extracted by layer separation [7], and the rectangular region centered at one particle. The major problem with Chen et al [22] is that tracking accuracy will decrease if mixed images have camera motion. Therefore, our proposed scheme applies motion compensation to both the prediction stage of particle filter (Section 2) and layer separation (Section 3) for tracking on mobile mixed images.…”

“…As to color vision, physicists generally support the trichromatic theory (Young-Helmholtz theory) [28] while psychologists often agree with opponent-process theory [29]. Thus, based on the particle filter with compensated motion model [23], particle weights are optimized using both RGB and [I, R-G, Y-B] color histograms on the mixed images in [22]. To improve tracking accuracy, the [I, R-G, Y-B] color histogram is computed for the intersection of the region indicated by the dynamic layer, extracted by layer separation [7], and the rectangular region centered at one particle.…”

“…Since few of previous trackers tackle the problem of reflection interference and multi-target tracking should discuss the areas of interactive multiple motion (IMM) model, data association, and state estimation in depth [24], this paper focuses on how to achieve robust single object tracking using multiple cue integration for mobile mixed images. This paper improves the work in [22] and its major contributions are stated as follows. (1) The proposed scheme improves tracking accuracy under the condition of reflections by combining co-inference [4] and maximum likelihood for visual cue integration.…”

“…Then each particle weight is optimized using maximum likelihood. One problem with Chen et al [22] is that tracking accuracy will decrease if videos have camera motion. This is mainly because the inaccurate static layer (i.e., reflectance image) makes edges of background and reflections contaminate the dynamic layer (i.e.…”

“…Tracks of separated layers are longer than those of the mixed images. Since the focus of tracking is the target but neither the transmitted scene nor the reflected scene, layer separation [13] also extracts the dynamic layer before single object tracking in [22]. Based on the framework of particle filter with compensated motion model [23], the correction stage reweights particles using RGB and [I, R-G, Y-B] color histograms of the mixed images [20].…”