A Shared Multi-particle Filter for Ball Position Estimation in RoboCup Small Size League Soccer Games

摘要

In the RoboCup Small Size League (SSL), teams of six robots which are autonomously controlled by a team AI play soccer games using an orange colored golf ball. In SSL games, an image processing system (SSL-Vision) receives images from four cameras which are installed 4 meters above the field surface, identifies the robots and ball positions. Because the SSL-Vision identifies the robots and the ball positions from all camera images, multiple position coordinates are computed in overlapping regions of the images. A shared multi-particle filter was developed to determine the ball position in the overlapping regions, and to minimize the discontinuity of the moving ball locus across the regions. The shared multi-particle filter consists of 4 particle filters and a particle binder estimates the position of the ball from the median point of a combined distribution. In an experimental setup, the shared multi-particle filter smoothly connected two ball locus vectors with a distance of 17 mm. The distances between each sample of the shared multi-particle filter were almost the same as both sets of data from the two cameras.