Unit detection from American football TV broadcast using multimodal content analysis

In this paper, a multimodal unit detection framework to detect and extract units, a novel concept towards event detection and extraction in sports TV broadcasts, is proposed. The proposed unit is defined to be a segment of a sports TV broadcast that describes a potentially interesting event, which possesses the potential of attracting the attention of the observer and satisfy his/her need of viewing the more interesting segments of the broadcast. A number of events that are considered as the unit target events are game impacting events such as score, missed score, penalties, and special game inserts, such as highlights and statistics clips. The proposed framework serves as an efficient data preprocessing procedure that can reduce the amount of data by ridding off the irrelevant data and prepare the remaining data in an efficient way for future event detection and extraction. Several experiments are conducted on various football games from different TV broadcasts, including college football and professional football. The experimental results demonstrate that the proposed framework effectively achieves the goal of data reduction, which is expected to increase the accuracy of event detection and extraction from the American football TV broadcast.