結合景深資訊與影像顏色特徵之平均位移物體追蹤演算法

Abstract

本論文提出了一套可用於複雜背景與移動場景下的即時三維物體追蹤演算法，影像追蹤演算法採用平均位移演算來計算目標物於空間中的位移，並利用主成分分析法，來估測出物體於三維空間中的大小與旋轉，因為加入了三維空間資訊，在複雜的背景下比一般的平均位移演算法具有更佳的追蹤效果，之後加入了Spatial-Color 的概念，使得目標模型有更精確的描述，在相似顏色障礙物靠近時，有一定的穩定效果。最後使用其演算法應用於機器人平台追蹤行人上，結果顯示有不錯的效能與穩定度。

In this thesis, a real-time tracking algorithm which can be used under complicated algorithm is based on Mean-Shift Algorithm. It combines principle component analysis to estimate the size and rotation of the target in 3-Dimensions. With the 3-D spatial information, this algorithm out-performs the traditional Mean-Shift. To enhance robustness, the concept of spatial-color is incorporated and thus giving better model to the target. When some object having the same color as the target approached, proposed algorithm is more robust, while traditional Mean-Shift might not work well. Experiments are presented using proposed algorithm to solve human-tracking problem on robots, which shows reliable and robust performance.