俯視型行人計數系統
No Thumbnail Available
Date
2008
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
本研究目的是發展出能計算區域範圍內經過的行人數之系統，利用攝影機模擬人工觀測的方式，取得所需的監測影像序列。本系統分為行人偵測和行人計數兩大部分，行人偵測主要工作為前景(foreground)擷取，行人計數的工作則為行人追蹤及計數。
在行人偵測中，最常遭遇的問題就是光影的變化及影像的雜訊，所以我們提出利用兩張連續影格的物件偵測方法，用較短時間差來減少光線影響，並使用較不受影像中雜訊影響的邊緣作為物件邊界，最後由邊緣及注意力區域將物件的邊界以Level Set方式封閉取出，使得影像中的行人區塊擷取出來。
在行人計數中，為了解決行人區塊彼此之間重疊合併或分裂以及不同攝影機架設角度下行人區塊形狀多變的問題，我們使用粒子濾波器(particle filter)進行行人的追蹤。粒子濾波器藉由同時提出多種假設並藉由統計算出運動模型來預測出移動物體的位置及形狀等狀態。本實驗中利用橢圓來代表粒子濾波器中的行人狀態，並使用橢圓的中心位置、大小、形狀、顏色等為特徵，量測出粒子與所追蹤行人區塊特徵值之間的相似程度(likelihood)，以更新每個粒子的權重值，再由計算粒子之期望值狀態，作為行人追蹤結果。之後我們將行人區塊依照其追蹤過程，定義成不同的行進狀態(state)，藉由狀態之間的轉換，進行行人數量統計。
由各種行人之間分裂合併測試，以及雨天、晴天、夜晚、和長時間高行人流量等各種環境實驗結果得知，本研究所提出之技術對於快速移動及物體大小改變顯著之行人，均有良好的追蹤效果。
In this thesis, we propose a bi-directional people counting system based on top-view video sequences. The system is divided into two sub-systems: people detection and people tracking. For people detection, we extract foreground object by two features of object: attention region and the boundary between objects and background. The attention region is generated by motion detection. Since people may stay at same location, our system should consider the stopping objects into attention region. Then we can use level set method to extract the objects. For people counting, we use particle filter to tracking objects to solve object merge-split problem. Each particle is represented by an ellipse. One object is tracking by a set of particles. The likelihood function of the particle weight is defined by location, color, and shape style. Using expect state to be output of our system. Then we use the tracking result to count number of people. Our system has been tested in different lighting conditions (e.g. weather, time, and environment) and using video sequences catching from different camera types (e.g. ordinary, and fish eye cameras) to show the robust of system.
In this thesis, we propose a bi-directional people counting system based on top-view video sequences. The system is divided into two sub-systems: people detection and people tracking. For people detection, we extract foreground object by two features of object: attention region and the boundary between objects and background. The attention region is generated by motion detection. Since people may stay at same location, our system should consider the stopping objects into attention region. Then we can use level set method to extract the objects. For people counting, we use particle filter to tracking objects to solve object merge-split problem. Each particle is represented by an ellipse. One object is tracking by a set of particles. The likelihood function of the particle weight is defined by location, color, and shape style. Using expect state to be output of our system. Then we use the tracking result to count number of people. Our system has been tested in different lighting conditions (e.g. weather, time, and environment) and using video sequences catching from different camera types (e.g. ordinary, and fish eye cameras) to show the robust of system.
Description
Keywords
行人計數, Level Set, 粒子濾波器