Target Tracking For MSL Soccer Robots

In the RoboCup Middle Size League competition, a soccer robot should be able to track multiple robots or obstacles by its own sensors, normally its omnidirectional vision system, with high accuracy. In this background, this dissertation mainly concerned with the multiple target tracking problems in such dynamic competition environments.This paper first presented an algorithm for target identification and localization, which is based on the imaging characteristic of the omnidirectional vision system and the specific appearance characteristics of targets. The proposed algorithm utilized a scan-line approach to determine the target positions. Experiments have shown it effectiveness and its noise characteristics are analyzed.After that, a constrained tracking filter for single target is proposed. This filter consists of a “Current” statistics model for target motion prediction, which is considered the most appropriate one after a comparison between various basic models. Also, by introducing the maximum acceleration constraint in the filter, it become less sensitive to the measurement noise.Finally, the algorithm for multiple target tracking is presented, which is base on the Joint Possibilities Data Association algorithm. Experiments have conducted to testify it effectiveness, and the results of the experiment have shown that this algorithm works well when targets are getting close and even if they collide with each others.