|Institution:||University of Missouri – Columbia|
|Full text PDF:||http://pqdtopen.proquest.com/#viewpdf?dispub=10157320|
Pedestrian detection and counting have important application in video surveillance for entrance monitoring, customer behavior analysis, and public service management. In this thesis, we propose an accurate, reliable and fast method for pedestrian detection and counting in video surveillance. To this end, we first develop an effective method for background modeling, subtraction, update, and shadow removal. To effectively differentiate person image patches from other background patches, we develop a head-shoulder classification and detection method. A foreground mask curve analysis method is to determine the possible position of persons, and then use a SVM (Support Vector Machine) classifier with HOG (Histogram of Oriented) feature and bag of words to detect the head-shoulder of people. Based on the foreground detection and head-shoulder classification at each frame, we develop a person counting algorithm in the temporal domain to analyze the frame-level classification results. Our experiments with real-world surveillance videos demonstrate the proposed method has achieved accurate and reliable pedestrian detection and counting.