Primena CBIR tehnika u biometrijskoj identifikaciji osoba na osnovu hoda

Abstract

Intenzivan razvoj informaciono-komunikacionih tehnologija otvorio je vrata primeni biometrijskih tehnologija u menadžmentu identiteta. Biometrijski modalitet koji ima veliki potencijal za primenu u praksi je ljudski hod. Njega odlikuju neinvazivnost i neintruzivnost. Ovakve osobine posebno pogoduju primeni u uslovima tehnologije prismotre. Zahvaljujući tome, ovaj biometrijski modalitet tokom prethodnih godina izaziva veliko interesovanje akademske zajednice. Ovo interesovanje rezultiralo je razvojem velikog broja pristupa za prepoznavanje osoba na osnovu hoda. Uprkos tome, primena biometrijskih tehnologija zasnovanih na ljudskom hodu u praksi i dalje zaostaje za dobro ustanovljenim modalitetima poput otiska prsta, lica ili glasa. Glavni razlog je nedostatak odgovarajućeg pristupa koji bi omogućio stabilnu primenu u realnim uslovima. Cilj ovog rada je predlog novog postupka za prepoznavanje osoba na osnovu hoda koji bi omogućio razvoj robusnog i pristupačnog biometrijskog sistema. Inicijalno, urađen je sveobuhvatan pregled oblasti i aktuelnih istraživanja na osnovu čega je predložen novi postupak. Predloženi postupak se zasniva na ideji da se sekvenca ljudskog hoda može predstaviti kao jedna nepomična 2D slika. Ovakav postupak omogućio bi da se za potrebe prepoznavanja primene generičke metode za pretragu slika na osnovu sadržaja. Na ovakav način problem bi bio prenet iz prostorno-vremenskog domena u prostorni domen, konkretno domen 2D nepomične slike, koji je poznat i u kome postoji veliki broj dokazanih rešenja. Za potrebe akvizicije, postupak se oslanja na novu tehnologiju iz oblasti interakcije čovek-računar, Microsoft Kinect. Na osnovu predloženog postupka razvijen je modularni laboratorijski prototip kao i okruženje za testiranje i evaluaciju. Naučna zasnovanost i opravdanost predloženog postupka proverena je nizom eksperimenata. Eksperimenti su organizovani na takav način da ispitaju različite faktore koji tokom primene postupka mogu uticati na konačne performanse u prepoznavanju. Na osnovu dobijenih rezultata može se zaključiti da predloženi postupak odlilkuje visok stepen robusnosti kao i visoka preciznost u prepoznavanju...

Intense progress of information and communications technology enabled application of biometric technology in identity management. Human gait, as a biometric modality, has great potential for practical application. This is due to its noninvasive and nonintrusive nature. Surveillance technology is especially fertile ground for recognition based on human gait. These facts caused spike in academic interest for this biometric modality. This in turn resulted in development of large number of different approaches to human gait recognition. Nevertheless, practical application of biometric technology based on human gait still trails those well established modalities such as fingerprint, face or voice. Main reason for this is lacking of such approach that would enable stable use in realistic conditions. Goal of this paper is to propose a new approach for human gait recognition that would result in robust and affordable biometric system. Initially, a comprehensive review of research area and existing research was done that served as a base for the proposition of new approach. This new approach is based on the idea that human gait sequence can be represented as a single 2D still image. Using images would open the possibility of applying Content Based Image Retrieval (CBIR) techniques for the purpose of final recognition. This procedure shifts the problem form spatio-temporal towards spatial domain, specifically the space of 2D still image that is well researched and familiar. For acquisition purposes approach relies on new human-computer interaction technology, Microsoft Kinect. As proof of concept, a modular laboratory prototype was developed as well as environment for testing and evaluation. Foundation of the proposed approach was tested through a series of experiments. Empirical evaluation was performed in such a manner to investigate the influence of different contributing factors to system performance. Based on retrieved results a conclusion is reached that the proposed approach is highly robust and achieves high recognition rates...