부산대학교 도서관

Automated approaches to analyze sports video content have been heavily explored in the last few decades to develop more informative and effective solutions for replay detection, shot classification, key-events detection, and summarization. Shot transition detection and classification are commonly applied to perform temporal segmentation for video content analysis. Accurate shot classification is an indispensable requirement to precisely detect the key-events and generate more informative summaries of the sports videos. The current state-of-the-art have several limitations, i.e., use of inflexible game-specific rule-based approaches, high computational cost, dependency on editing effects, game structure, and camera variations, etc. In this paper, we propose an effective decision tree architecture for shot classification of field sports videos to address the aforementioned issues. For this purpose, we employ the combination of low-, mid-, and high-level features to develop an interpretable and computationally efficient decision tree framework for shot classification. Rule-based induction is applied to create various rules using the decision tree to classify the video shots into long, medium, close-up, and out-of-field shots. One of the significant contributions of the proposed work is to find the most reliable rules that are least unpredictable for shot classification. The proposed shot classification method is robust to variations in camera, illumination conditions, game structure, video length, sports genre, broadcasters, etc. Performance of our method is evaluated on YouTube dataset of three different genre of sports that is diverse in terms of length, quantity, broadcasters, camera variations, editing effects and illumination conditions. The proposed method provides superior shot classification performance and achieves an average improvement of 6.9% in precision and 9.1% in recall as compared to contemporary methods under above-mentioned limitations. [ABSTRACT FROM AUTHOR]