Automatic Detection of Slam Dunk Events in Basketball Game Video

• Main Authors: Ben-Tai Miao, 苗本泰
• Other Authors: Hui-Huang Hsu
• Format: Others
• Language: zh-TW
• Published: 2009
• Online Access: http://ndltd.ncl.edu.tw/handle/21782832357569254293

碩士 === 淡江大學 === 資訊工程學系碩士班 === 97 === Because of the spread of multi-media broadcast and fascinated Basketball sport, people who want to watch live or replay program is more than before but not all of them have time to enjoy the whole game show. Although there are top ten selections from the one week program, there are still some good actions that are not selected. Therefore, we hope we can construct a system that can pick out the “Slam-Dunk” events from the basketball game videos for those who are addicted to basketball or has a job to pick out the “Slam-Dunk” events from the thousands of videos. This thesis can be divided into four parts. At first, we should extract the location of score board. Firstly, we make those which edge pixels have been appeared mostly in the whole video into a gradient map. Though the operations of closing and opening, we eliminate the noises and gather the connected parts. Next, we find the location of score on the scoreboard. According to the gradient map we did it before, we take down the changing pixel within it so that we can get the changing information thought time to time. Then, we select those candidates according to their height-width ratio to find out the locations of scores. The third part is score identification. We focus on the region of score location and extract the explicit of score digit. After identifying scores, we only focus on those score events which equal to two. The final part of the thesis is identifying whether the two-point events are slam-dunk or not. We pick up the two-point events. During the events, we use our method to select the right view frame and find the hoop firstly then count the skin pixel near the hoop position in order to find the hands grabbing the hoop. If the skin pixels are greater than our threshold, we claim it as a slam-dunk event. The contribution of this research is we save much time for those who are addicted to slam-dunk actions and those who need to find out them. And the more, we prove that using low level feature in image processing still can reach the high level semantic event.