Research On Highlights Detection In Soccer Videos

The increasing mass of digital videos has boosted the need for video contents via networks, and mobile devices particularly. To reduce the data volume and save time for audience, highlights are selected manually from videos. However, the ever increasing demand has drawn researcher’s attention to the automatic detection of highlights. In this paper, we focus on soccer video event detection in three aspects: the structural analysis of videos, the extraction of features which contain important semantic clues and the fusion of multiple features as well as event type discrimination.Firstly, according to the analysis of current structural units, we utilize shots as the physical segment and Play-Breaks as the event clip. As for shot boundary detection, we combine both color and edge related features, with extra preprocessing and postprocessing techniques. Using semantic shots of view types and replay scenes, we propose a rule-based Play-Break generation algorithm.Secondly, with regard to the pattern features of Goal Scored event and Goal Attempt event, we extract 9 low and mid level features as the descriptors of Play-Break event clips. Apart from penalty area detection, we propose the concept of shot focus, and provide the extraction method. Shot focus is an identification of Goal related events, and as a bonus, the mid field area of soccer playfield could also be detected. Besides, we further explore the Corner Kick event. Such event segments exceed the boundaries of Play-Breaks. Hence we propose K-P shot pair as the event unit. Features such as corner area and playfield distribution are extracted.Lastly, for the low and mid level features from multiple sources, we propose to utilize Multiple Kernel Learning for feature fusion and metric learning. Combined with SVM, the MKL-SVM model is finally used to decide event types such as Goal Scored, Goal Attempt and Corner Kick. Experiments tested on various soccer videos demonstrate the effectiveness of our method.