Theme-based Sports News Video Retrieval Research

Since the 2000s, the demands for interesting sports news are increasing. To enhance the efficiency of sports news retrieval system, more and more researchers add the theme annotation of sports news into process of retrieval. This paper deeply studies about some key technology of sports news retrieval based on theme, such as shot boundary detection, anchor shot detection, feature of shot extract and theme annotation. The main achievements lie in the following aspects:(1) This paper proposes a method of improved space-time module feature to detect shot boundary. Space-time models are adopted to detect the local continuity of shot. Meanwhile, this paper presents a mechanism of fusion based on graph theory. This mechanism can combine the modules which are used to describe the same object in order to avoid the false detection or undetected errors of shot boundary(2) This paper presents a new method of detecting anchor to segment the news units. First of all, according to the characteristics of modest action and fixed motion area of the anchor, motion information is adopted to choose candidate anchor shot. Then, considering the feature that anchor appears frequently in the screen, this paper use accumulated similarity to detect the final anchor shot from candidate anchor shots. Finally, the video between two sequential anchor shots is outputted as a news unit.(3) Considering the problems that traditional SIFT points cannot present motion information of object in the video, this paper presents an improved SIFT point to describe feature of shot. Utilizing the characteristic of space-time interested points to choose the SIFT points which present the information of motion object. Then, k-means algorithm is adopted to cluster the chosen SIFT points. Finally, the SIFT points which can describe theme of sports news are outputted as feature of shots.(4) In the step of annotation, for the inability of traditional Hidden Markov Model to use multi-dimensional feature to annotate, this paper adopts Embedded Hidden Markov Model to annotate the theme of news units. The main HMM is adopted to record the random transfer process of the shots of news unit, while the sub-HMM is used to record that of the features of each shot. Considering the problems that HMM output model can easily fall into the trap of local extremum, this paper presents a new training model. Meanwhile, the QPSO is used to optimize the process of training to gain the global optimal resultsThe experimental results show that these methods proposed in the paper can solve the existing problems efficiently.