利用動態攝影機實現人形追蹤與人臉偵測技術

Abstract

本論文提出一套基於適應性動態攝影機控制技術之智慧型人形追蹤與臉部偵測系統，在我們的系統架構下，整個系統可以區分成兩個部分。第一個部分是以影像為基礎之偵測模組，其中包含了人形和人臉的偵測；第二個部分則是針對pan-tilt-zoom攝影機所開發的控制模組。首先，我們利用人體的模型來辨識影像中存在的人形，再用膚色資訊和connected component找出人臉的位置。利用所找到人臉的資訊及相關特徵，我們完成了一個可用來判斷人臉是否清晰可見的神經網路系統。因此，本論文所提出之技術不但可將被追蹤的人保持在攝影機可視範圍的中間；當移動中人員的人臉不夠清楚或是尺寸小於我們所設定臨界值的時候，系統會自動調整攝影機焦距的倍率，以便得到清晰可辨識的人臉影像。

This thesis presents a study of human tracking and face detection on a pan-tilt-zoom camera. In our approach the system is divided into two parts: detection module including human and face detection, tracking module including pan-tilt-zoom camera control. We model a human for human recognition and combine HSI skin color with connected components for face detection. HSI skin color is implemented by using multilayer neural network trained with back-propagation learning algorithms. We use the recognizable face index to describe the definition of detected faces. Active camera will track a human and keep it in the FOV (Field of View) camera. The controls in camera zooming will be automatically started when the detected face of moving humans is not clear or the face index is smaller than a threshold value. We also apply local motion vectors to increasing the capability of our tracking system.