在群體人數估計應用中使用不同特徵與回歸方法之分析比較;Estimating Number of People in Groups using Different Features and Regression Method: A Survey

NCUIR > College of Electrical Engineering & Computer Science > Graduate Institute of Computer Science and Information Engineering > Electronic Thesis & Dissertation > Item 987654321/54399

Please use this identifier to cite or link to this item: http://ir.lib.ncu.edu.tw/handle/987654321/54399

Title:	在群體人數估計應用中使用不同特徵與回歸方法之分析比較;Estimating Number of People in Groups using Different Features and Regression Method: A Survey
Authors:	張世佳;Chang,Shih-chia
Contributors:	資訊工程研究所
Keywords:	電腦視覺;人群估計;智慧型監控系統;computer vision;people counting;intelligent surveillance system
Date:	2012-07-17
Issue Date:	2012-09-11 18:49:59 (UTC+8)
Publisher:	國立中央大學
Abstract:	由於電腦視覺的技術已經漸漸發展到較為成熟的階段，以及電腦計算效能的提升，愈來愈多人投入了智慧型視訊監控系統的研究，目前架設在各處的監控式攝影機如果沒有人為的操作，很難在緊急的情況下發揮效用，因此更加需要發展智慧型監控系統，使其能夠在無人的狀態也能發揮監控的效果，例如在商場中，若能使用智慧型監控系統來進行人潮流量的估算，便能夠提供相當有用的訊息給商場的負責人，在人群行為分析方面，若能夠利用智慧型監控系統來分析人群的行為，即時發現異常的行人流量，在安全上也能夠提供完善的資訊，因此智慧型監控的發展已經愈來愈受到重視。本論文的目的是估計複雜環境下移動中的人群內行人的數量，主要分析人群中可以偵測到的各種特徵對於人數估計的結果比較。在複雜的環境，我們利用連續影像相減法來取得移動中的人群位置，對移動中的人群做不同的特徵擷取，分別是角點的偵測、SURF點的偵測以及頭部肩膀部分的偵測，並將擷取到特徵數量用來建成不同的特徵組合，使用類神經的單層感知機以及Support Vector Regression兩種回歸分析方法來建成估測人數的回歸模型。在進行人群估測的時，我們同時考慮了分群與不分群的方式，不分群的方式就是將整張圖片擷取到的特徵數量一起訓練，而分群的方式則是增加了密度的資訊，因為畫面中的人群分佈距離鏡頭的遠近並不相同，所以我們預估不同人群所包含的特徵數量雖然相同，但是前景的大小卻是不同的，因此人群中的人數應該也會有所差異，參考了密度資訊以後，我們一樣將其特徵數量建立回歸模型，並用來估測人數。在實驗的部分，我們使用PETS2009的資料集，從中挑選出四個適合的場景來做人群數量的估算，並且比較各種特徵組合與回歸方法估測的結果，總結來說，使用愈多的特徵種類來估算人數能夠達到較小的誤差，詳細的討論與分析將會在論文當中完整描述。Computer vision technology has gradually developed into a more mature stage. Today, with the prevalence of surveillance cameras, automated and intelligent surveillance functions are desired. People counting and crowd analysis has become an important topic in intelligent surveillance applications. In the thesis, we survey the methods of estimating the number of people in moving crowds via regression in a complex environment. We analyze the effects of features and regression methods one the performance of estimation.First, we use frame differencing to get the approximate location of moving crowds. Afterwards, we detect three different types of features from these locations. The features are the number of detected corner features, SURF features and head-shoulder regions. Finally, we use perceptron and Support Vector Regression to build regression models to estimate the number of people in the crowds. Also, we consider the way of clustering the foreground connect-components into clusters. The distances between the clusters and the camera are different. Since clusters nearer to the camera would have more foreground pixels and result in more detected feature points, the estimation would be affected. Therefore, we also consider density information into the feature combination and build regression models to estimate the clustered crowds.We use the public PETS 2009 dataset to perform the experiments. We pick four appropriate views for estimating and compare each feature combination and regression model. Overall speaking, adding more types of features in the feature combination results in smaller estimation errors. Detailed analysis and discussions on the performance are explained in this thesis.
Appears in Collections:	[Graduate Institute of Computer Science and Information Engineering] Electronic Thesis & Dissertation

Files in This Item:

File	Description	Size	Format
index.html		0Kb	HTML	699	View/Open

社群 sharing

Loading...