Moving Object Detection Based On Type-2Fuzzy Set And Probabilistic Graphical Model

Moving object detection is widely used in intelligent video surveillance,human-computer interaction, military application fields etc. It's also the foundation oftarget identification, classification and behavior analysis and understanding. It's one of thebasic research work in the field of computer vision. There are a lot of confounding factorsin moving object detection such as complex backgrounds(shaking trees and lawn, changingcloud and waving water), illumination changing(slow and sudden) and something else(shadow removal, filling holes and etc.). These factors lead the challenge to implementaccurate and effective moving object detection. The paper does deeply a lot of work onmoving object detection in complex dynamic scenes and keeps the focus on the accuracyand robustness of foreground objects. The main research contents are as follows:1). Systematically doing a comprehensive analysis to gaussian mixture model and itsimproved algorithms. We find that pixels process do not strictly satisfy the gaussiandistribution since the noise in complex scenes. It's more obvious to misjudge thebackground/foreground when using the traditional gaussian mixture model in dynamicscenes. The paper introduces type-2fuzzy sets theory (T2FSs) based on GMM. It uses theprimary membership function to handle the uncertainty of pixel distribution in GMM, thenthe uncertainty of the primary membership function is considered by the secondarymembership function. The model can describe the pixel distribution accurately, andexperiments show that the method can effectively filter the foreground noise and get a veryclean background.2). Foreground objects holes are obvious in low contrast conditions when use themodel in1), so a method using texture information, gradient information or similar isconsidered. But the feature extraction in existing block-based models is complex and slow. The paper proposes a combined model between block-level and pixel-level gaussianmixture background model. The discrete cosine transform (DCT) is done to each sub-blockof video frame. A pseudo image is constructed by using the most important4DCTcoefficients of each block. The pseudo image contains rich spatial features extraction, andthe model based on the pseudo image is very efficient in the shadow depression and holesremoval. A simple probabilistic way combines the result here and in1) which can enhancethe accuracy and completeness of the foreground object.3). Though the spatial information is considered in2), it's still a pseudo-pixel patternwhich separates the relationship between global and local information in video frame. So,Markov random field (MRF) model is introduced to model the relationship. The pixel labelis decided by both its neighborhood and the global energy function. The paper models thebackground prior knowledge getting from the1) and2) and the spatial consistencybetween the pixels with MRF. Using the local features association of Markov property inthe MRF to refine the contour of the foreground objective and meanwhile, the graph cutsmodel is introduced to minimize the energy function generated by MRF.