本論文旨在設計並實現多人員追蹤辨識系統，並在魚眼影像上進行分析，為了避免魚眼影像經過校正處理後會有部份影像被裁切掉，故我們不對魚眼影像做任何的校正處理。我們將魚眼攝影機架設於天花板上且視角向下拍攝，影像中的人員物件會隨著與魚眼影像中心點相對應的位置不同，而產生不同的樣貌姿態，為了能夠有效的偵測到魚眼影像中變形的人員物件，我們使用近年來迅速發展的深度學習演算法來達成人員追蹤辨識的目標，透過深度學習的技術讓機器從龐大的資料庫中，其中亦包含變形的人員圖片，學習並辨識圖片的特徵，其目的為解決魚眼影像所導致的變形問題且透過大量的資料訓練可強健網路對人員的姿態變化的適應能力，不因人員的姿態變化而需另行設計系統進行辨識。 本文所實現之演算法主要分成兩大部份，第一部份為利用YOLOv2深度網路架構偵測魚眼影像中的人員物件並進行定位，第二部份為使用Deep SORT架構分析在連續影像中所偵測到的人員物件是否為同一人員並進行追蹤，並使用AlignedReID網路架構將已辨識出的人員再進行辨識，進而確認其人員的身份，整個系統的設計主要是藉由深度學習的技術所完成，進而強健此系統的功能。 ;This thesis attempts to design and implement a multiple human tracking and identification system that is applicable to omnidirectional surveillance by a fisheye camera. In this work, a top-view fisheye camera is mounted on the ceiling. Image distortion correction is not performed on each image frame because the human objects far from the image center might be cropped after lens undistortion. The appearance of human varies depending on the position in an image, and thus it is hard to detect and recognize human by traditional approaches. Accordingly, the deep learning technique developed rapidly in recent years is employed for achieving efficient human tracking and recognition. By learning from a big amount of training images, the computer will have the ability of extracting features, detecting and identifying human. The main algorithm of the proposed deep neural network includes two stages of networks. The first part is a YOLO-based deep architecture for detecting and locating human objects. The second part combines location and appearance information to track an identical person. Through the entire process of tracking a person, the identification will be kept as the same by using a so-called AlignedReID network. From the experimental results, the efficiency and robustness of the proposed algorithm have been verified.
