Video Image Segmentation and Object Detection Using Markov Random Field Model

摘要

In this dissertation, the problem of video object detection has been addressed. Initially this is accomplished by the existing method of temporal segmentation. It has been observed that the Video Object Plane (VOP) generated by temporal segmentation has a strong limitation in the sense that for slow moving video object it exhibits either poor performance or fails. Therefore, the problem of object detection is addressed in case of slow moving video objects and fast moving video objects as well. The object is detected while integrating the spatial segmentation as well as temporal segmentation. In order to take care of the temporal pixel distribution in to account for spatial segmentation of frames, the spatial segmentation of frames has been formulated in spatio-temporal framework. A compound MRF model is proposed to model the video sequence. This model takes care of the spatial and the temporal distributions as well. Besides taking in to account the pixel distributions in temporal directions, compound MRF models have been proposed to model the edges in the temporal direction. This model has been named as edgebased model. Further more the differences in the successive images have been modeled by MRF and this is called as the change based model. This change based model enhanced the performance of the proposed scheme. The spatial segmentation problem is formulated as a pixel labeling problem in spatio-temporal framework. The pixel labels estimation problem is formulated using Maximum a posteriori (MAP) criterion. The segmentation is achieved in supervised mode where we have selected the model parameters in a trial and error basis. The MAP estimates of the labels have been obtained by a proposed Hybrid Algorithm is devised by integrating that global as well as local convergent criterion. Temporal segmentation of frames have been obtained where we do not assume to have the availability of reference frame. The spatial and temporal segmentation have been integrated to obtain the Video Object Plane (VOP) and hence object detection In order to reduce the computational burden an evolutionary approach based scheme has been proposed. In this scheme the first frame is segmented and segmentation of other frames are obtained using the segmentation of the first frame. The computational burden is much less as compared to the previous proposed scheme. Entropy based adaptive thresholding scheme is proposed to enhance the accuracy of temporal segmentation. The object detection is achieved by integrating spatial as well as the improved temporal segmentation results.