Font Size: a A A

Research On Sports Video Content Analysis Technology

Posted on:2006-03-14Degree:DoctorType:Dissertation
Country:ChinaCandidate:F WangFull Text:PDF
GTID:1118360185995707Subject:Computer application technology
Abstract/Summary:PDF Full Text Request
With the rapid increase of the amount of digital videos, the problem how to find useful information has become much urgent. This thesis focuses on content-based video analysis and retrieval technology in sports video domain. The aim is to process, analyze and understand video content with computer to construct index and structure for facilitating user's access, which is provided with important academic appeals and commercial potentials.The core of sport video analysis technology is to extract semantic events and their relationships from raw video data. It is a complex and challenging problem. The essential difficulty is the gap between low-level features processed by computer and high-level semantics understood by human. Our main research efforts focus on the problem, discussing the principles for integrating domain knowledge to pattern analysis in video which is a kind of multi-dimensional signal containing both spatial and temporal information. In the thesis, we firstly introduce the method to automatically extract a kind of specific semantic events, i.e. highlights. Then for more general semantic events, we discuss rule-based and statistical-based methods to sports video analysis.Sports video highlights represent important or interesting video segments, which are not specified exactly. Existing method exploit visual or aural features to build subjective model for highlight detection. For emotion and feeling are hard to be described by low-level features, the performance is not very effective. Given the observation that highlights are often followed with slow-motion replay segments, we propose a replay-based method for highlight extraction. Our method combines color features and camera motion analysis into shot boundary detection, replay logo recognition and motion pattern match for highlight extraction. Experimental results indicate that our method is very effective, with 92% precision rate and 98% recall rate, which are higher than about 75% rates by using subjective models.Motivated by principles of natural language processing, we talk about rule-based methods for sports video analysis and propose a video parsing system by using grammar. Firstly, we segments video into elementary shots. Then through event detection, shots are annotated with...
Keywords/Search Tags:Video Analysis, Video Retrieval, Sports Video, Semantic Event Detection, Rule-Based, Statistical-Based, Highlight, Dynamic Bayesian Network, Multimodal
PDF Full Text Request
Related items