Digital characterization using computer vision in real-time

Robert Ravnik (2009) Digital characterization using computer vision in real-time. Prešeren awards for students.

Preview

PDF Download (3217Kb)

Abstract

Digital signage in combination with computer vision opens new possibilities for out-of-home advertisement. By determining certain characteristics of observers (e.g. gender, age) the information presented to them may be adjusted accord- ing to their expected preferred interest. To denote the process of detecting, characterizing and responding to the observer we introduce the term \digital characterization" as our system significantly reaches over the existing so called \digital signage" systems that only presents information without detecting and characterizing the observers. Digital characterization in real time has become possible due to the development of the image analysis methods and the devel- opment of the widely available computer equipment. A computer system for digital characterization is described. It is an intelligent system for displaying selected visual information on a computer screen. The system tracks and char- acterizes the viewers by analyzing the images of their faces taken by a camera attached to the screen, using the computer vision methods in real time. It also performs logging and analysis of recorded data. Main components of the sys- tem are the player and the main server. The player performs characterization of the viewers which can be used for adjusting the information on the screen. The main server performs selection of the presented information and manip- ulates the data. Detection and tracking of several observers is performed in real time. In face characterization we use the Principal Component Analysis (PCA) and the data mining environment Orange. Characterization of the ob- server's gender from their face is specially treated. The following methods of machine learning are applied and inter-compared in the Orange environment: naive Bayes, K-nearest neighbors, classification tree and random forest. A part of FERET face image library is used as learning and test sets. Selected methods are implemented for real time application on a PC. Gender of an ob- server can be determined in 17,9 ms with 83,3% reliability using random forest classifier. A web application for managing of the system and for generating reports is developed, including statistical analysis and visualization of data. The system is designed for standard computer and video equipment. For this reason it is practically applicable wherever we want to adapt the information to the viewer: in education, medical institutions, marketing, etc

Actions (login required)