The Research Of Content-based Feature Extraction And Retrieval Method For Surveillance Video

With the progress of technology and the improvement of social security consciousness, more and more surveillance cameras are widely used in various aspects of social life. The data produced by all of surveillance cameras in the world in a second can reach trillions level. So, how to retrieval the target from mass surveillance videos effectively and quickly has become a difficult problem. This thesis studies how to make use of the characteristics of the image to retrieve monitor video and how to utilize the LIRE image retrieval framework to retrieve key frames of surveillance videos. Finally, this thesis puts forward three kinds of feature extraction method.Knowing that different characteristics of the image in HSV can lead to the different sensitivity of people’s eyes from the human visual characteristics, we quantity the color to get color features by normalized method. And at the same time we use the AC coefficient after the DCT transformation to extract image texture information. By mixing color and texture features together, we can get a novel color texture feature. Not only can this mixed color texture feature improve the image retrieval in recall and precision ratio, but also the monitor video retrieval subjective effect is good.After view of the traditional binary pattern extraction of texture feature, we find it is too simple and not suitable for the issue of content-based image retrieval. Thus we put forward a more suitable method for content-based retrieval called three- value-pattern feature which based on the binary mode characteristics. And the experimental results show that the effect of feature of content-based image retrieval is much better than the traditional method in recall and precision ratio.Due to precision rate of mixing color texture feature on the content retrieval is not high. The thesis proposes a fuzzy color texture feature extraction method in this thesis. In this method we use 24 kinds of fuzzy color and 16 kinds of texture feature(two types of point, four types of straight line, and the four types of corners, five types of cross and a type of flat area). And the experimental results show that the fuzzy color texture feature is better than CEDD(Color and Edge Directivity Descriptor) in the retrieval precision.The thesis proposes a kind of target feature extraction method based on the interested region owing to people pay more attention to the human and vehicles in the practical application of monitoring video retrieval. The final result indicates that the method presented in this thesis can achieve the goal of retrieve the interested area accurately.