於單一入口處，允許行人裝扮更換之影像再辨識系統之研究

Abstract

隨著科技普及，許多場所都有架設監控攝影機以保障社會安全，並且陸續有越來越多自動化監控系統的相關研究以減少人力負擔及時間的花費，但在攝影機無法拍攝之死角處，如公共廁所、辦公室等，經常被有心人士進行犯罪後用來更換裝扮以混淆監控系統的判斷，造成社會安全的疑慮。本論文針對以上問題，設計一套於單一入口處，允許行人裝扮更換之影像再辨識的系統，主要目的在連結行人自身關聯性，即使在裝扮發生更換時仍可進行再辨識，使監控系統建立完整的行人資訊，達到實際應用之功能。 在此，本論文的貢獻有以下兩點：第一，本論文提出一篩選候選人機制，結合單視野測量法計算出之身高特徵、步態能量圖提取出之步態特徵及皮膚偵測判斷出之頭部特徵，挑選出可能性較高之候選人，以提高行人裝扮更換後之再辨識的成功率；第二，針對貢獻一所挑選出之候選人進一步提取其行人色彩及紋理特徵並加入高斯混和模型來抑制背景的影響，之後本論文藉由改良的相似度演算法，在候選人中找出最終可能性最高之行人。

As technology advances, many monitor cameras are set to maintain social security anywhere. In order to reduce the cost of manpower, lots of automatic monitoring systems are researched. However, there are many blind angles, the outlaw will use these blind angles to crime and change their clothes to confuse the automatic monitoring systems. In this thesis, we design a pedestrian re-identification online system allowed pedestrian to change clothes under the single entrance circumstance. There are two contributions of this thesis. First, this thesis proposes the candidate screening mechanism which used the single view metrology to calculate the height of pedestrian, gait energy image to extract the gait feature, and skin detection to determine the head feature. Thus, the result of pedestrian re-identification will be more robust when the pedestrian is changed the clothes. Finally, the candidates are extracted the color and texture feature which is added the Gaussian mixture model. Then, this thesis improves the similarity algorithm to match the most like pedestrian.