生醫教師追蹤與視訊錄影

Abstract

本論文開發一套即時的生醫教師追蹤與視訊錄影系統，能夠在室內環境中對特定目標教師以一廣角相機與一左右轉動上下傾斜與放大縮小 (PTZ, Pan-Tilt-Zoom) 相機來進行追蹤。 行人追蹤已經有許多的生活應用，例如:老年人看顧、線上即時會議、住宅安全監控。在此追蹤系統中，我們採用由偵測來追蹤的方法以粒子濾波的框架來實現。 首先我們必須指定欲追蹤目標，接著以粒子濾波來模擬目標位置的分布。我們訓練一卷積類神經網路來估計目標影像是否為人頭之機率，還有以背景相減法來做前景偵測，且以色彩直方統計計算候選與目標之相似度。最終以估計得最高機率位置為當前追蹤之目標，再迭代估計下一幀之位置。

In this thesis, we develop a nearly real-time biomedical teacher tracking and video recording system to track a biomedical teacher in the indoor scene by one wide-angle camera and one PTZ (Pan-Tilt-Zoom) camera. Human tracking has many applications such as eldercare, security surveillance, and online meeting. In this human tracking system, we employ tracking-by-detection in particle filter framework to track the target. We have to specify which person to track first, and model the location of the target as a state distribution by particle filter. Moreover, we train a convolutional neural network as a head classifier to estimate the probability of human head, the motion detector with background subtraction, and color histogram is used to obtain the similarity between candidate and the target.