Robust detection of moving objects in video sequences through rough set theory framework

Detection of moving objects in the presence of challenging background situations like swaying vegetation, rippling water, camera jitter etc., is known to be a difficult task. Background subtraction is considered to be better than the other approaches in terms of robustness. Its success primarily depends on the proper choice of background model(s) associated with every pixel for its foreground/background labeling. In this work, we have adopted rough-set theoretic measures to embed the spatial similarity around a neighborhood as a model for the pixel. Basic histon and its associated measure Basic Histon Roughness Index (BHRI) have been reported in the literature. It was applied to still image segmentation with impressive performance. Its adoption in video sequences for foreground/background labeling is proposed herein. We extended the histon concept to a 3D histon, which considers the intensities across the color planes in a combined manner, instead of considering independent color planes. Further, we also incorporated fuzziness into the 3D HRI measure. The labeling decision is based on Bhattacharyya distance between the model HRI and the corresponding measure in the current frame. Adoption of rough set theoretic concept into moving object segmentation is nontrivial, as the model updating requires careful consideration so that the pixels associated with gradually changing background or dynamic background are labeled as background and at the same time, slow moving objects are never adopted into the background model. A novel background model update strategy proposed herein takes these into consideration and also eliminates the need of having exclusive ideal background frame initially.