利用粒子濾波機制與動態檢查表來達到多視角的人臉追蹤

Abstract

在視訊監控的應用中，人通常是最主要的觀察目標，因此人臉的偵測與追蹤是一個重要且具有實用價值的研究方向。本論文提出了一個多視角的人臉追蹤方法，主要的追蹤機制是採用粒子濾波為基礎，並且修改了其中的轉移性函式與相似性函式來符合我們的應用需求。除此之外，我們提出動態檢查表針對不同的人臉姿勢記錄相對應的歷史資訊做為比對目標的依據，藉此來達到多視角的人臉追蹤並且同時利用動態更新的比對基準使追蹤更能應付追蹤環境的變化。最後，我們設計了校正機制使得在追蹤發生失敗時，能擴大搜尋的範圍來找尋目標，並在找到目標的時候，再一次回覆追蹤的狀態。實驗的結果顯示我們所提出的方法對於人臉在不同側臉的角度變化、人臉大小的縮放與光線的改變化都有不錯的追蹤效果，同時對於人臉遭到遮蔽時能切換到校正狀態來搜尋人臉，並且當人臉再度出現在畫面時回到追蹤狀態來達到持繼追蹤人臉的目的。Automatic and robust detection and tracking of faces is essential in the surveillance application, especially when the main targets are about people. In this research, we propose a novel multi-pose face tracking method. Our method is based on the Condensation algorithm with modified transition and likelihood function. A dynamic update structure, called Checklist, is designed to help identify the current face post from the target history and for storing the latest face. Furthermore, template-matching-based algorithm is used to recover lost face during tracking. Experimental results show that the proposed algorithm can deal with the change of poses and illumination. It is also able to extract target face information even when the tracking is lost for whatever reason.