An effective multi-object detection approach

In snooker sport, there are most 22 balls on the table, which have different colors. In television relay, we usually need to show the “Snooker” (or location of balls on table) in 3D view. It requires accurate location of balls. In this paper, we propose an approach to locate all the balls automatically. Our approach includes training stage and location stage. In the training stage, we extract the joint color histogram and range of valid area for different color balls. Then we compute average color vector and obtain their range for different color balls. In the location stage, joint color histograms are used to get candidate ball region. Then false regions are excluded by area. Finally true color ball is confirmed by the distance of average color vector between the candidate ball and trained values. Our approach is tested using 21000 balls in 1000 training images and 20800 balls in 1000 testing images, 92.86% balls can be located correctly in training images and 91.23% balls can be located exactly in testing images.