Research Of Multi-Moving-Object Segmentation Technology Based On SOFM In MPEG Compressed Domain

With the rapid development of multimedia technologies and emergence of 2G video coding, the technology of video motional object segmentation, significant part of multimedia application, will be applied to such fields as video coding, retrieval, browsing, smart video surveillance, virtual reality, computer vision, pattern recognition and multimedia communication based on object, widely in the future. However, no general method to video motional object segmentation is accepted and presently, scenes can be divided into many semantic objects and coded by content-based representation and scalability that will be playing an important role in 2G technologies of video coding and segmentation.Video segmentation is to divide an image or sequence into some areas according to the specifications in order to extract some significant entities, called video object in digital video processing. Video object segmentation is currently an active area of research and some success has been reported in the field. But most of the approaches are performed in the pixel domain. The technology of video motional object steps from that of still image segmentation and it includes two categories mainly: interframe-based segmentation and intraframe-based segmentation. Firstly, the original image data is reduced to segmentation, which can be implemented, by low-pass filters and media filters, etc; secondly, features of image data, i.e. color component, texture, motion, frame difference and semantic information will be extracted; thirdly, the determination to segmentation with some uniform standard is made based on the features. The main advantage of that is the precision of segmentation.But due to the fact that most of video data are stored in compressed form, and in order to extract the objects we have to decode them first. In other words, more amount of time is required to extract the objects. For the reason above, more and more researchers are studying the method for extracting the objects directly from the compressed video sequences, and this field is becoming a hot topic in resent years. Video objects segmentation based on motion vectors and DC coefficients are taken more considered, and system based on motion vectors is adopted by our paper.Artificial neural network (ANN) is an active and crossing discipline and research on it is of significance, which is composed of a multitude of neurons connected mutually to simulate the properties of human brain. Its information processing is achieved by the function among different neurons connected and the storage of information and knowledge is represented by the physical contacts of network components. ANN has been adopted to classify and cluster image and video data.MPEG compressed video provides one motion vector for each macro block of size 16×16 pixels. Initially, the motion vectors are scaled appropriately to make them independent of frame type. This is accomplished by dividing the motion vectors by the difference between the corresponding frame number and the reference frame number (in display order). Then, the motion vectors are rounded to the respective nearest integers. In the case of bidirectional predicted macro blocks, reliable motion information is obtained either from the forward or backward motion vector depending upon which of the reference frames (I/P) is closer. If backward prediction is chosen, the sign of the motion vector is reversed after normalization. Finally, these vectors are coded.Based on the study of the previous algorithms in compressed domain video segmentation,ANN and the structure of MPEG, I develop a fast, automatic system to extract the multi-object from compressed MPEG video. The method takes the compressed video sequence as input and the motion vectors are decoded from the interceded (P and B) frames. Then motion vectors are accumulated over a few frames from the reliable macro blocks. Then the temporally accumulated motion vectors are subjected to median filtering to get the dense motion field. How to divide the vectors is difficult due to insufficiency of data. This problem is overcome by Kohonen algorithm, which is an iterative technique that alternately estimates and refines the segmentation and motion estimation. SOFM is adopted to cluster and segment videos in this paper, which is anti-noise, adaptive and self-organizing ANN and can learn without human surveillance and the training data will also be applied to other compressed video sequences. The dense motion vectors are given as input to the Self-Organizing Feature Maps (SOFM) models, and they are divided into several sorts. Each sort belongs to one object, and we can extract them correctly now.Numbers of experiments proved that the system can adapt to most of videos, and it is real-time. But still some problems such as the difficulty of two objects have the similar motion vectors segmentation, which are the main work in the future.