Event Detection of Broadcast Baseball Videos

This paper presents an effective and efficient event detection system for broadcast baseball videos. It integrates midlevel cues including scoreboard information and shot transition patterns into event classification rules. First, a simple scoreboard detection and recognition scheme is developed to extract the game status from videos. Then, a shot transition classifier is designed to obtain the shot transition patterns, which contains several novel schemes including adaptive playfield segmentation, pitch shot detection, field shot detection, as well as infield/outfield classification. The extracted midlevel cues are used to develop an event classifier based on a Bayesian Belief Network. The network is with low complexity because the number of these cues used is small, which not only improves the generalization performance of the event classifier but also reduces system complexity as well as training efforts. Using the inference results of the network, we further derive a set of classification rules to identify baseball events. The set of rules is stored in a look-up table such that the classification is only a simple table look-up operation. The proposed approach is very simple and computational efficient. More importantly, the simulation results indicate that it identifies ten significant baseball events with 95% of precision rate and 92% of recall rate, which is very promising.