Research On Soccer Video Index Structures And Retrieval Algorithms

With the development of video searching technology, the content based video retrieval has attracted more and more attention from researchers in recent years. General content based image retrieval system uses only the low-level features, which is insufficient to cover all the essential information, thus leading to the results that are far from satisfying the users. To solve the problem, this essay studies how to combine the low-level features and high-level semantic features to modify current index structure and build an efficient soccer video index structure.The video retrieval system is composed of three main parts: feature extraction, index building and retrieval algorithm designing. For feature extraction part, according to the shortcoming of existing soccer game image view classifying ways, a new fluctuation times detecting method is proposed to recognize the soccer image view type. The method applies a window sliding in the center of the image from left to right, calculating the field ratio of the slide window each time and recording the field ratio change times between the upper and the lower of the in-field view threshold. The paper also proposes a novel high-low position method to classify the soccer playfield zone. After getting the field area by using Gaussian Mixture Model and pre-processing the field edge, the high-low position method could be able to judge the playfield zone by the coordinate's information of only the left edge point and right edge point. For the index building part, the existing ERVQ index structure is ingeniously modified to put in the view type and playfield zone semantic information, creating a tree like structure to partition the index posting list according to the semantic information. For the retrieval algorithm part, the paper proposes a novel voting algorithm to quickly select candidates and design an algorithm to sort the candidates by color features.Experimental result shows that the fluctuation times detecting view type and high-low position classifying playfield zone methods are very promising, outperforming two state-of-art approaches. And the index structure designed in this paper is better than current ERVQ index in both searching time and returned results. To optimize search experience further, our future work will focus on importing more semantic features of video and designing a distributed index structure for big data.