應用於主動式攝影機上的權重式重取樣粒子濾波器的人形追蹤

Abstract

我們提出了權重式重取樣粒子濾波器演算法，並應用於主動式攝影機上做人形追蹤。本系統主要有三部分，分別是人形偵測，人形追蹤和攝影機的控制。在人形偵測方面使用編碼比對去得到人形區域，而粒子濾波器會對每張輸入的影片估測出人形的位置。我們在重取樣中選擇具有高權重的粒子，因為這能讓追蹤特徵更精確。此外比例-積分-微分控制器(PID controller)用來控制主動式攝影機，藉由最小化介於影片中心和物體位置的誤差，並轉換此誤差為pan-tilt速度來驅動攝影機移動讓被追蹤的人保持在影片的可視範圍內。在追蹤過程中，影像的強度和人的特徵也會隨之變化。因此高斯混和模型(GMM)會隨著時間來更新人的特徵模型。至於短暫的遮蔽問題可透過特徵相似度和重新取樣粒子來解決。而粒子濾波器能估測出每張輸入影像裡人形的位置，因此可以平滑地驅動主動式攝影機。

We proposed a weighted resampling algorithm for particle filter and applied for human tracking on active camera. The system consists of three major parts which are human detection, human tracking, and camera control. The codebook matching algorithm is used to extract human region in human detection system, and the particle filter algorithm is estimating the position of the human in every input image. We select particles with high weighting value in resampling, because it will give higher accurate tracking features. Moreover, a proportional–integral–derivative controller (PID controller) controls the active camera by minimizing difference between center of image and the object’s position obtained from particle filter, also convert the position difference into pan-tilt speed to drive the active camera and keep the human in the field of view (FOV) camera. The intensity of image may change over tracking, and so do the human features. Therefore, the Gaussian mixture model (GMM) algorithm is used to update the human feature model overtime. As regards, the temporal occlusion problem is solved by feature similarity and the resampling particles. Also, the particle filter is estimating the position of human in every input frames, thus the active camera will drive smoothly.