Modelling Video Frames for Object Extraction Using Spatial Correlation

Object extraction forms a critical part of the object-based video processing. However, most of the techniques available concentrate only on surveillance and tracking. Normal video sequence does not have steady background and hence these techniques cannot be applied to them. In our work, we propose an elegant method to model background and foreground based on histogram data. We use 2D continuous wavelet transform to spatially localize object and create object mask to approximate silhouette. With available histogram for object pixels and background pixels, we obtain probability density function by normalizing the area under histogram. In order to retain smoothness in our density function, we use curve-fitting techniques to approximate the probability density function.