| With more and more installations of digital camera in cities, explosive information is generated which is essentially important for the increasing requirements of the social security. Its usage for surveillance and police is well-known. Feature extract in the video sequence is one of the most crucial part of intelligent visual surveillance system. Considering the model deficits of the state-of-the-art algorithms, our research is focused in the following three aspects:Firstly, considering the diversity of the multiple feature extraction algorithms, as well as their model redundancy, we propose a feature-level information fusion framework based on linear dimensionality reduction. The combined multiple feature vectors are projected into a low-dimensional feature space when the redundancy and nuissance information are maximally suppressed. Meanwhile, the model complexity is reduced.Secondly, the video trajectory is considered in the module of feature extraction. The trajectory implies a special distribution of feature space. Also, reducing the dimensionality of feature space of video is highly necessary. Thus, in the current theoretical framework of linear dimensionality reduction, we integrate the distributional information of feature space by a graph, by which two kinds of classical algorithms are yielded. Another innovative point in this chapter is a new fusion algorithm on the feature level: merging two graphs.The last technical chapter includes the research on how to maintain the integrity of an image in the feature extraction model. Here we directly adopt a matrix-based representation of features, which is successful by using an iterative solution to the projection matrices.Using the MATLAB platform we implement two classical applications of visual surveillance: crowd counting and video/image recognition. We also conduct a comparative study between classical methods and proposed methods, which indicate that our proposed methods yield better performance. |
PDF Full Text Request