本論文設計實現一個具有兩個攝影機的立體視覺系統來探討目標物體的距離量測，並且將其應用於FIRA人形機器人競賽的競走項目。在目標物體的距離量測上，本論文探討攝影機參數的校正、特定目標的標定、特徵點的擷取、特徵點的匹配以及距離的量測，並且應用類神經網路做誤差修正來得到一個更正確的距離資訊，同時利用立體視覺的深度擷取特性以及特徵點的位置判別來讓機器人自主判斷其目前所在的位置，進而完成競走比賽的策略。實驗結果驗證本論文所提之類神經立體視覺演算法確實可以獲得較精確的距離資訊。 A stereo vision system with two cameras is implemented in this thesis. Based on the stereo vision system, a distance measurement of a target object is investigated and applied to the Sprint event of FIRA HuroCup. In the distance measurement of the target object, these topics of the camera calibration, specific target selection, feature point extraction, feature point matching, and distance measurement are discussed. A neural network is used to obtain more correct measurement data. The depth capture of stereo vision and the location judgment of feature points are used to let the robot can autonomously determine its current location. Furthermore, the strategy for the Spring event of FIRA HuroCup is completed. Some experimental results and comparison are presented to illustrate the implemented method can indeed obtain more accurate distance information.
