Visual Tracking Based on Color Kernel Densities of Spatial Awareness

We propose a kernel-density based scheme that incorporates the object colors with their spatial relevance to track the object in a video sequence. The object is modeled by the color probability density function across a set of pixel regions on the object, partitioned in terms of the base shapes such as the concentric annuli or polygons at the object centre. The probability density of the object is derived by applying the kernel density estimator region-wise to the pixels within such partitioned areas. This proposed object representation enables the independent processing of the color features while at the same time making the implicit use of location information without having to involve additional model parameters. Weighting factors are also introduced to differentiate the significance of the relative physical locations when measuring the similarity of two probability density functions, and this facilitates the tracking of the more robust object features. The located object is then finalized for its boundary deformation by demanding a neighborhood similarity in colors at the object pixels near the boundary. Our experimental results showed this method is effective at tracking a non-rigid object on a moving background.