英文摘要 | With the growing requirement for security, intelligent visual surveillance is now seeing increasing applications. As the low-level problem in intelligent visual surveillance, moving object detection is the basis for the subsequent steps, such as object tracking and object classification. Therefore, it is the key issue for the application of visual surveillance. In this thesis, we concentrate on moving object detection based on background modeling. The main contributions of this work are as follows: 1.We present a novel method for shadow removal using Markov Random Field (MRF). In our method, we first construct the shadow model based on hierarchical information. At the pixel level, we use the GMM to model the behavior of cast shadow for every pixel in the HSV color space. At the global level, we will exploit the statistical feature of shadow in the whole scene over several consecutive frames to make a preclassifer accurate and adaptive to the change of shadow. Then, based on the shadow model, an MRF model is constructed for shadow removal. The main novelty of our method is that, although our method is chroma based method, we can make the preclassifier accurate and adaptive to the change of shadow by using the statistical feature of shadow at the global level. 2.The KDE (Kernel Density Estimation)-GMM (Gaussian Mixture Model) Hybrid model (KGHM) is proposed to overcome some problems of KDE and GMM for background modeling. For the KDE method, one has to save many samples for accurate density estimation, and the selection of bandwidth is also very important for estimation. For the GMM method, the number of Gaussian components should be set manually. Moreover, both methods do not fuse spatial information to deal with highly dynamic scenes. The KGHM can overcome the problems described above, when applied in the background modeling. In the KGHM, the spatial information is fused in two ways: first, KGHM is constructed in grid (bin) level; second, the kernel function is used to model the dependencies between a sample and its neighboring bins. Our method does not need to store samples, as the coordinates of every pixel which are described by kernel function don’t change over time. Moreover, we use the Gaussian merging and deleting method to determine the number of Gaussain components in the GMM part, dynamically. 3.Based on the KGHM, we propose a moving object detection method which can deal with shadow and dynamic background, simultaneously. We first constru... |
修改评论