使用馬可夫鏈蒙地卡羅方法之多方位行人偵測
No Thumbnail Available
Date
2010
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
本論文研究多方位行人偵測的技術，攝影機可以不受架設位置與觀測角度的限制偵測行人。為了要掌握行人於影像中呈現各種不同的型態，我們提出一個多視角(multiple-view)的單位球(unit sphere)來描述行人，稱此單位球為viewsphere，它是由多個巢狀球面所組成，每一層球面均勻分佈許多視點(viewpoints)。我們將一3D行人模型置於球體中心，然後將行人模型投影至每一視點所屬的影像平面，因此可以取得各種不同觀測角度的行人外觀，稱為model views。
本研究首先建立一3D行人模型，此行人模型是由行人頭部和肩膀所組成的上半身，因為上半身的輪廓形成”Ω”形狀，為行人獨有的特徵，即使在擁擠的人群中，這個輪廓也不易消失。利用此輪廓資訊再搭配頭髮的髮色、臉部的膚色和所占區域的面積比資訊，可以於影像中找出行人的位置，即使行人發生遮蔽情況也可成功標示出行人的位置和計算行人的數目。由於行人狀態的解空間很龐大，我們利用馬可夫鏈蒙地卡羅(Markov Chain Monte Carlo)的方法，在解空間中連續取樣，計算行人於影像中的後驗機率 (posterior probability) 分佈，再根據後驗機率分佈，決定出行人最佳狀態。由於馬可夫鏈蒙地卡羅收斂速度慢，因此我們設計三種不同的取樣策略，提升建立後驗機率的效率。
實驗時，在不同場景架設不同高度和不同攝影角度的攝影機，測試本研究所提出的技術。結果證明，可適應各種角度的監視影像，且行人發生遮蔽的情況下，也能正確找出行人位置。
This paper presents a technique for multi-view pedestrian detection. The camera can be mounted anywhere with any viewing direction. We use a multiple-view unit sphere, called the viewsphere, to represent pedestrian. The viewsphere forms a 2D manifold of viewing directions (i.e., viewpoints), which are unidistributed over the spherical surface of the viewsphere. The pedestrian detection problem is formulated as maximum posterior estimation. Due to the high complexity of the solution space, we explore the solution space using a Markov Chain Monte Carlo (MCMC) sampling method. Three kinds of proposal distribution are proposed to further improve the efficiency of MCMC.
This paper presents a technique for multi-view pedestrian detection. The camera can be mounted anywhere with any viewing direction. We use a multiple-view unit sphere, called the viewsphere, to represent pedestrian. The viewsphere forms a 2D manifold of viewing directions (i.e., viewpoints), which are unidistributed over the spherical surface of the viewsphere. The pedestrian detection problem is formulated as maximum posterior estimation. Due to the high complexity of the solution space, we explore the solution space using a Markov Chain Monte Carlo (MCMC) sampling method. Three kinds of proposal distribution are proposed to further improve the efficiency of MCMC.
Description
Keywords
多方位, 行人偵測, 馬可夫鏈蒙地卡羅, 單位球體, 3D行人模型, 遮蔽, Multi-View, pedestrian detection, Markov Chain Monte Carlo, viewsphere, 3D human model, occlusion