Zaznavanje avtomobilov v slikovnih podatkih s pomočjo računalniškega vida

DEBENJAK, SEBASTJAN (author), Štruc, Vitomir (mentor)

V diplomskem delu se posvetimo problemu detekcije avtomobilov v slikah na način, da udejanjimo in ovrednotimo lasten detektor vozil. Detektor udejanjimo s pomočjo deskriptorjev histograma orientiranih gradientov (angl. Histogram Of Oriented Gradient – HOG) in razvrščevalnika podpornih vektorjev (angl. Support Vector Machine – SVM). Za namene primerjave delovanja razvitega detektorja udejanjimo še kaskadni razvrščevalnik Haar, primerjavo pa ovrednotimo s krivuljo natančnosti priklica (angl. precission-recall curve). V uvodnem delu najprej predstavimo sorodna dela, ki predstavljajo zgodovino razvoja metod za detekcijo vozil v slikovnih podatkih, ter grobo primerjavo med njimi. Opis nadaljujemo s poglavjem, ki predstavi teoretične osnove nekaterih nujno potrebnih delov za detekcijo objekta na sliki, v poglavju s teorijo pa se osredotočimo zgolj na postopke in metode učenja, ki so potem tudi uporabljeni pri implementaciji lastnega detektorja. V nadaljevanju diplomskega dela je najprej opisan postopek luščenja deskriptorjev s histogramom orientiranih gradientov v kombinaciji z razvrščevalnikom podpornih vektorjev. Druga metoda pa predstavlja implementacijo in detekcijo s pomočjo kaskadne metode Haar. V drugem delu podajamo postopek izvedbe detektorja. Opišemo in prikažemo pridobitev podatkovne zbirke, uporabo postopka HOG za pridobitev značilk, učenje detektorja SVM in uporabo detektorja. Poleg tega prikažemo, kako različni parametri vplivajo na uspešnost detekcije in odvisnost predobdelave vzorcev na detekcijo. V zaključku naloge pa podajamo primerjavo med metodami v uspešnosti in hitrosti detekcije objekta.

• Language: Slovenian
• Keywords: zaznavanje in detekcija avtomobilov, deskriptor, razvrščevalnik SVM, HOG, Haar, podatkovna zbirka, OpenCV
• Work type: Bachelor thesis/paper
• Organization: FE - Faculty of Electrical Engineering
• Year: 2019
• PID: 20.500.12556/RUL-109297
• Publication date in RUL: 29.08.2019

Secondary language

• Language: English
• Title: Detecting cars in image data using computer vision

Abstract

This diploma paper focuses on the problem of vehicle detection in images with a development and evaluation of a new vehicle detector. The vehicle detector uses the descriptors of a histogram of oriented gradient (HOG) and the support vector machine (SVM) classifier. To compare its functioning to the functioning of a working detector, the Haar cascade classifier is used and the results are evaluated with the precision-recall curve. The introduction starts off with a presentation of similar works that detail the history of developing methods for detecting vehicles in image data and a rough comparison between them. This description is followed by a chapter on theoretical bases of some integral parts for detecting an object in an image. The theoretical chapter only focuses on the processes and methods of learning, which are later used in the implementation of the new detector. In the next part of the diploma paper, the process of extracting descriptors via the histogram of oriented gradient in combination with the support vector machine classifier is described. The other implemented method is the implementation and detection with the help of the Haar cascade method. The second part focuses on the detector’s developmental process. It includes a description and demonstration of the acquisition of the database, the use of the HOG process for the acquisition of attributes, the learning process of the SVM detector, and the use of the detector. This is followed by a description of the influence of different parameters on successful detection and the effect of pre-treatment on detection. The paper then concludes with a comparison of methods based on the effectiveness and the speed of object recognition.

• Keywords: car recognition and detection, extraction of features, SVM descriptor, Histogram of Oriented Gradient, Haar, cars dataset, OpenCV