本文提出一種基於直方圖更新的混合粒子濾波法，藉由結合AdaBoost分類器與傳統粒子濾波法，改善傳統粒子濾波法之目標追蹤效率。首先，利用AdaBoost分類器能夠快速分辨目標物與背景的特性，來偵測目標物在影像中的位置，並周期性地更新目標物的參考色彩直方圖，以供粒子濾波器使用；接著，由粒子濾波器對目標物作預測、量測的動作，以找出最有可能是目標物的位置；最後，由訓練後的倒傳遞類神經網路，將目標物的影像資訊轉換為真實的座標資訊，以達到追蹤之目的。本文進行了在不同粒子數與不同光照環境下的L型追蹤，以及目標物消失後再出現之不同追蹤實驗，並與傳統粒子濾波法比較。由實驗結果顯示，本文所提出之方法在以上實驗能達到良好的追蹤效果，且效果優於傳統粒子濾波法。

This paper presents a histogram updating-based hybrid particle filter for object tracking robot applications. In the proposed method, a well-trained AdaBoost classifier periodically detects the object’s location in an image and generates a new reference color histogram for the object. The new reference color histogram is then used by a traditional particle filter to perform the prediction, measurement, and sorting operations, in order to produce the moving object’s most possible location in an image at the current time-step. The obtained location is then passed to well-trained back propagation neural networks to generate the moving object’s polar coordinate, including the distance and angle, relative to the follower robot. Eventually, the obtained polar coordinate is used to create speed and rotation commands, which make the follower robot move towards the object, in order to keep it in the camera’s central field of view. Different experiments, including the L-shape tracking under different number of particles and different lighting conditions, and the tracking of an object that temporally disappears from the camera’s field of view, are conducted using both the presented hybrid particle filter and traditional particle filter. Experimental results show that the presented hybrid particle filter outperforms the traditional particle filter.