隨著科技的進步，觀眾對於運動轉播影片所能提供額外的娛樂功能也益增，以棒球轉播來說，從單純的在轉播比賽中主播口述並邀請球評講解比賽中的細節，演進至影像中提供球速資訊，甚至更進階需要擁有專業知識的資訊如棒球中的配球理論，這些一般大眾較少能夠接觸到的球賽資訊，也是觀眾較為感興趣的部分。因此本論文以棒球投手球種辨識為目標，整合棒球軌跡追蹤以及球路辨識。此外，並提出以階層式boosting為基礎之球路辨識演算法，取出棒球飛行軌跡特徵後，以其Adaboost分類器延伸之應用於多類別的SAMME演算法選出有效特徵，而弱分類器的設計我們採用以多類別貝式分類法，最後再以這些被賦予不同權重的特徵代表的弱分類器組合成一球路辨識之強分類器。實驗結果顯示，對於多種不同的球場影片，我們的球種辨識正確率平均可以接近80%。 With advances in technology, the audiences expect additional entertainments of sport broadcasts. To baseball broadcasting, the pure broadcast game only with anchor oral and ball comment discuss the details of the game, go on appear the speed information on screen, even provide more advanced information that audiences have a few opportunities to understand, like pitching tactics theory, this part was attracted to the audiences, too. In this paper, we integrate the pitch recognition with trajectory tracking. Moreover, we propose the hierarchical boosting based pitch. The strong learner adopts the SAMME algorithm that is extended from Adaboost, the weak learner is designed based on the multiclass Bayesian classifier. Our experimental results show that accuracy of pitch recognition could be reached near 80%.
