基於模板更新之孿生網路的三百六十度視訊等角立方體投影之行人追蹤;People Tracking Based on Siamese Network with Template Update for EAC Format of 360-degree Videos

NCUIR > College of Electrical Engineering & Computer Science > Graduate Institute of Communication Engineering > Electronic Thesis & Dissertation > Item 987654321/83790

Please use this identifier to cite or link to this item: http://ir.lib.ncu.edu.tw/handle/987654321/83790

Title:	基於模板更新之孿生網路的三百六十度視訊等角立方體投影之行人追蹤;People Tracking Based on Siamese Network with Template Update for EAC Format of 360-degree Videos
Authors:	戴鸛臻;Tai, Kuan-Chen
Contributors:	通訊工程學系
Keywords:	行人追蹤;360度視訊;等角立方體投影;孿生網路;FLD;貝氏分類器;people tracking;360-degree videos;equi-angular cubemap (EAC);Siamese neural network;Fisher linear discriminant (FLD);Bayes classifier
Date:	2020-07-15
Issue Date:	2020-09-02 17:06:31 (UTC+8)
Publisher:	國立中央大學
Abstract:	在360度視訊中，等角立方體投影(equi-angular cubemap projection, EAC)屬於立方體投影(cubemap projection, CMP)的變體，等角立方體投影相較於立方體投影幾何形變程度較小，在追蹤問題上較不易衍生錯誤。然而等角立方體投影的影像仍具有相鄰面內容不連續的特性，且仍有不均勻幾何失真，導致現有追蹤方案在等角立方體投影的影像中準確性嚴重下降。因此，本論文針對等角立方體投影的360度視訊，提出基於孿生網路的行人追蹤方案，以卷積神經網路(convolutional neural network)對目標模板與目前畫面之搜索視窗提取特徵，並比對特徵以追蹤目標。在影像不連續的問題上，本文使用面拼貼 (face stitching)措施，使追蹤器能於連續的影像內容進行追蹤，同時避免造成更多幾何形變。因應不均勻幾何失真，基於孿生網路由當前畫面計算的分數圖(score map)，來預測更新模板(template update)的時機，使用FLD (Fisher’s linear discriminate)將分數圖降維，並計算分數圖之平均值與標準差作為三種特徵，再通過貝氏分類器(Bayes classifier)決定是否更新模板。實驗結果顯示，本論文所提出之面拼貼與模板更新方案有效提升SiamFC追蹤準確率。;Variants of cubemap projection format (CMP) such as equi-angular cubemap (EAC) of 360-degree videos has less geometric deformation, which may reduce tracking error. However, accuracy and speed of most existing trackers degrade seriously in the face of content discontinuity and non-uniform geometric deformation in EAC formats of 360-degree videos. Thus, this paper proposes a Siamese network based people tracking scheme for 360-degree videos using EAC format. The tracker extracts features from the target template and the search window of the current frame by a convolutional neural network, and compare features to predict the bounding box of target. To be robust against the content discontinuity between inconsistent adjacent faces of EAC images, this paper proposes an efficient face stitching scheme such that the tracker keeps tracking across adjacent faces and avoids raising geometric deformation simultaneously. By referring to the score map generated by Siamese networks, the proposed pre-trained Bayes classifier based mechanism of template update determines the right timing of update. The input feature vector of Bayes classifier includes the data that generated by dimensionality reduction from score map using Fisher linear discriminant (FLD), the mean of the score map and the standard deviation of the score map. Experimental results show that the proposed face stitching scheme and the mechanism of template update effectively improve the tracking accuracy of SiamFC.
Appears in Collections:	[Graduate Institute of Communication Engineering] Electronic Thesis & Dissertation

Files in This Item:

File	Description	Size	Format
index.html		0Kb	HTML	148	View/Open

社群 sharing

Loading...