Research On Video Object Segmentation And Representation Based On Motion Feature Analysis

As the use of video becomes increasingly popular, and wide spread through, providing means for fast, automated, and effective techniques to represent video based on its content, such as objects and meanings, are important topics of research. Usually, video databases are very large so that it is difficult to implement in real-time applications such as transfer, processing, storage, representation, retrieval, especially in object-based video storage and multimedia mobile communication. Therefore, effective video compression, coding and representation would significantly shorten the transform bandwidth and reduce the costs of video retrieval and surveillance.The most important problem of video representation is video analysis based on object segmentation and feature extraction. The goal of this thesis is to provide stable object-based video segmentation and features representation methods. The content includes representing the history and current status of video object segmentation, comparing the performances of different video object segmentation algorithms, delineating structure of representation of object-based video representation, and present the our algorithms.The quality of video affects the output of video segmentation and features greatly. To alleviate the influence of noise and low contrast to segmentation, multi-scale morphological methods are used. Features of noise in video are estimated by analyzing the feature images of different scales extracted by open and close operators of morphological reconstruction. According to the features of noise, noise reduction can be implemented by decreasing the value of feature of noise in small scale space. On the contrary, local contrast of image can be enhanced by increase the value of feature in large scale. The detailed procedure of implementation is presented in this thesis.We presented a novel object segmentation algorithm based on analysis of motion and change features in wavelet domaind. The algorithm can obtain more information of object motion edges by analysis, detection and fusion of different scale, i.e. from low scale to high scale. To achieve higher integrity of object, signal-noise-ration of motion image with low scale is improved according to the motion feature of multi-image on the basis of independent propriety of noise distribution. However, the motion of different objects are different, an effective control strategy for motion information is presented by the estimation of motion to assure the integrity of object segmentation of video series. Experiment results show this algorithm is enough effective and timesaving to be used in real time application.Object tracking is the other key to video object segmentation and features representation. Object tracking method based on motion prediction is presented on the basis of continuity of object motion. The features of object motion can be calculated by object tracking according to minimum distant method, which maps object motion of previous frame to current one and compute the positions of motion object in feature space. Meantime, background is segmented by background regression to improve the accuracy of segmentation.Most estimations of object motion are based on cross-central-biased of motion vector. Besides this, we found another property of object motion: direction of probability distribution of motion vectors. No-full-symmetric double cross pattern search strategy is presented on this property. The method is timesaving because the search strategy is compact support. Theoretical inductions and experimental results show that our method can improve search speed greatly whether for complex large motion series or for simple small motion ones. The effectiveness of this search strategy is also verified by the experiment of estimation of object motion by fill techniques.In a word, the thesis discussed the basic features of object and presented video representation method based on senior features of object. Video content is difficult to represent only by finite basis features of object because video usually contains complex information. However, features of object are related and object track, motion vector, time delay, position in space can be fixed on by object segmentation, object tracking and object estimation etc. All the advanced features are useful and used in our algorithm. Although we presented some new methods to solve the problems in video segmentation and representation which are effective confirmed by theory and experiment, the generality of these methods should be improved further. To facilitate the usage of video, the related research should be more deeply.