標題: 利用多攝影機在人群中決定人物之位置與身高
People localization and height determination for dense crowds using multiple cameras
作者: 黃星陸
Huang, Hsing-Lu
莊仁輝
Chuang, Jen-Hui
多媒體工程研究所
關鍵字: 消失點;線取樣;多層參考平面;定位;平均前景覆蓋率;等間隔樣本點;遮蔽;多攝影機;Vanishing point;Line samples;Multiple reference planes;Localization;Average foreground coverage rate;Uniform sampling of ground plane;Occlusion;Multiple cameras
公開日期: 2010
摘要: 近幾年來以多攝影機進行多人物定位與追蹤的研究越來越受到重視,其中大部分的方法都需要大量的計算才能夠處理嚴重遮蔽的問題,也因此往往需要倚賴特殊硬體才能達成即時的定位與追蹤。不同於這些研究,本論文主要特色在於以取樣過的前景樣本,投影於地面以限縮人物可能出現之區域,再利用人物所在區域會被較多投影的前景樣本覆蓋之特性,設計一隨機演算法能於限縮的區域中快速地尋找出人物的數目及位置。由於所需的搜尋空間以及次數都大量的被減低,使得所需的計算量也大幅的被減少,因此我們的方法能夠提供即時的人物定位結果。此外我們利用了人物於空間中佔有一定的體積之觀念,以等間隔的地面取樣來排除因人物間彼此遮蔽所產生的錯誤偵測。進而達成在嚴重遮蔽情況下仍能正確地偵測出人物所在的位置,並提供準確的定位結果。
In recent years, researchers have been paying much attention on people tracking and localization using multiple cameras. Most of methods require a large number of computations to cope with serious occlusions, and need to rely on special hardware to achieve real-time locating and tracking. Unlike these studies, we use two dimension line samplings of foregrounds to restrict regions of possible locations of people. According to the nature of occupancy constraint that possible locations of people should be covered by more projected foreground pixels, we propose a random algorithm to efficiently find the locations and number of people. Our methods can provide real-time location results because of the smaller of searching space from line sampling and less computing time due to the randomness nature. Besides, we exploit the concept of people volume so that uniformly sampling on the ground plane can prevent error from occlusions. Experimental results show that our approach provides real-time and accurate people localization results under serious occlusions.
URI: http://140.113.39.130/cdrfb3/record/nctu/#GT079757533
http://hdl.handle.net/11536/46072
Appears in Collections:Thesis


Files in This Item:

  1. 753301.pdf

If it is a zip file, please download the file and unzip it, then open index.html in a browser to view the full text content.