Visual descriptors for Visual Place Recognition

By

TABLE OF CONTENTS

CHAPTER 2: BACKGROUND AND LITERATURE REVIEW1

Introduction1

Place recognition using Features2

Robotics and Place Recognition3

Place Recognition methods4

SIFT Descriptor7

Scale-space extrema detection8

Key points Localization10

Orientation assignment12

Key point Descriptor12

Why SIFT Features considered in this system?13

Bag of Visual Words14

CENTRIST Descriptor16

Spatial Weighting19

SURF20

Repeatable Angle Computation21

Image Descriptor Concatenation21

Color histograms22

Edge orientation histogram22

Edge extraction23

Matching23

Gradient histogram24

Gradient computation24

Spatial/orientation binning24

Spatial Semantics25

Place Classification32

Place Recognition32

Place Categorization38

Properties40

Semantic Mapping SCENARIO43

Place Classification44

REFERENCES46

CHAPTER 2: BACKGROUND AND LITERATURE REVIEW

Introduction

The accurate place recognition is a challenging task and it is extremely important that self-localization is enforced precisely. It contains some cases that illumination changing, some objects being removed or shifted around and/or people affecting the appearance of the scene [indoor place recognition]. The camera movement for the visual impaired or the mobile robot for the same place may get some difficulties of changing environment like

The image of the same place changed because illumination has been changed because changing of the light for the room and so on.

The image of the same place changed because of moving some objects or obstacles get in the way

The image of the same place changed because of change the side view of the camera so input space is huge, and lot of images need to be analysed , and recognized online to be suitable for the robot for their navigation or localization process.

The topic is widely researched, and a huge of research has been proposed, the most used is incremental learning approaches for constructing the geometrical map, or the environment representation, online [indoor place recognition]. The main and basic method used for the place recognition task is to encode some region of the place scene as a descriptor that can be stored in a dataset, in this chapter, a detail description of some important descriptors that have been used in place recognition techniques including this work.

Place recognition using Features

Features are distinct properties or pieces selected from the visual scene, used to know or recognize the visual scene, these features as interested points can be detected or extracted in many ways, each way gives a type of features differs from the other. In Place recognition, many types of feature detectors have been used to recognize the visual place scene including objects in front of the camera sensor. The best type is the one that not affected by the diversity of the environmental impact, like illumination, rotation, scaling, transformation, etc. Also some features are very important for recognition of landmarks, while the navigation process. The strong and best features affect of the localization for Visual SLAM system. VSLAM will be explained in details.

Features can be divided into two types, local and global. The global features are useful to recognize objects shapes or general form of the object scene, for example, using moment function to know the orientation of the object as a global feature. This type of features is sometimes used with the local features in CBIR, and it is difficult to use it in localization for mobile robotics. For this reason, the uses of local features are very common in navigation, due ...