Semantic Labeling of Soccer Video

2. Segmentation and Classification: To do segmentation, the video stream is first divided into relatively static parts and active parts. For static parts, motion features are ignored and key frames are saved. Every active part is further segmented into active sub-parts according to 4 view types (defined in Section 2). In Classification stage, motion features are used to classify (label) segments with Support Vector Machines [5]. As traditional shot segmentation may not produce video segments that possess one-to-one correspondence to semantic views, we present an integrated segmentation and classification approach to label soccer video into semantic units in this paper. In our system, each P frame is divided to a 6 by 4 blocks with color and motion features extracted on both block and frame levels. First, a threshold is used to divide the video stream into relatively static parts and active parts. Then every active part is segmented into subparts according to 4 view types and the motion features are used to classify segments with Support Vector Machines. Finally, static parts are merged with classified active subparts to form labeled segments. Four 10-minute test clips from the World Cup 2002 are used to evaluate our system resulting in a promising classification rate of 79.8%. 3. Post-Processing: static parts are merged with adjacent active sub-parts to form semantic segments with semantic labels.