多相机计算机视觉下的机器人SLAM与导航技术

slam

需积分: 38 183 浏览量更新于2024-07-19 3 收藏 24.29MB PDF 举报

身份认证购VIP最低享 7 折!

领优惠券(最高得80元）

"多相机SLAM在移动机器人导航中的应用" 移动机器人SLAM（Simultaneous Localization And Mapping，同时定位与建图）是机器人领域中的核心技术之一，它允许机器人在未知环境中自主探索并构建环境地图的同时，精确地估计自身位置。多相机SLAM则进一步提升了这一技术的精度和鲁棒性，通过利用多个摄像头捕捉的不同视角信息，可以更有效地解决定位和建图问题。本研究由Gerardo Carrera Mendoza博士在帝国理工学院计算机系进行，并在Dr. Andrew Davison的指导下完成。论文详细阐述了如何利用多摄像机系统实现移动机器人的精确定位和建图，展示了多传感器融合在SLAM中的潜力。多相机系统的优势在于，它可以提供更丰富的视觉信息，包括深度、立体视差以及更广阔的视野。这些信息有助于提高机器人对周围环境的理解，尤其是在复杂或动态环境下，多相机的冗余可以增强系统的稳定性和可靠性。例如，当一个摄像头的视线被遮挡时，其他摄像头仍能提供必要的视觉数据。在SLAM算法中，通常会涉及特征检测、匹配和数据关联等步骤。多相机系统中的这些步骤需要进行同步处理，以确保不同摄像头捕获的画面之间的一致性。此外，还可能涉及到几何校正和传感器标定，以减少由于镜头畸变、光照变化等因素引入的误差。论文可能涵盖了以下关键点： 1. **特征提取与匹配**：在每帧图像中检测出稳定的视觉特征，如SIFT、SURF或ORB，然后进行特征之间的匹配，以确定不同相机间的对应关系。 2. **数据融合与优化**：将来自多个摄像头的数据整合到一个统一的地图模型中，通常通过概率框架如扩展卡尔曼滤波（EKF）或粒子滤波来实现。同时，利用 Bundle Adjustment 等方法对整个SLAM过程进行后端优化，以提高地图质量和定位精度。 3. **定位与建图**：通过多视图几何方法，如立体视觉或结构化光，估算机器人的位姿轨迹，并同时构建环境的三维模型。 4. **鲁棒性与实时性**：探讨如何在面临光照变化、动态物体、传感器噪声等问题时保持系统的稳健性，以及如何设计高效的算法以满足实时性需求。 5. **实验与评估**：可能包括在实验室环境和真实世界场景下的实验，通过比较单相机与多相机SLAM的性能，展示多相机系统的优越性。这篇论文深入探讨了多相机SLAM技术，为移动机器人在未知环境中的自主导航提供了重要的理论基础和技术支持。通过多角度信息融合，提高了机器人定位和建图的精度，对于推动机器人领域的实际应用具有重要意义。

资源详情

资源推荐

Introduction

1.2.1 Consumer Robots Already on the Market

The market for practical mobile robots is still small but is growing in importance. Since

the year ∼ 2000 we have perceived an increasing development of home-robotic appli-

cations, mostly in the ﬂoor cleaning area. In this section some of the most successful

home-robotic products are analysed. However, for more a extensive discussion, the

reader should refer to [109] .

In 2001 Electrolux lunched the ﬁrst cleaning robot for mass production, the “Trilo-

bite”. This robot uses a ring of sonars covering 180 degrees to determine the perimeter

of the room and to avoid obstacles. It also has the capability of recharging itself in a

station base when the battery is about to deplete. In Figure 1.1 can be seen a compari-

son of some robotic vacuums.

In 2003, Karcher releases the Robocleaner RC3000. This device has a random

operating mode which is helped by tactile sensors to gradually cover the area and to

avoid obstacles. Its main advantage is the usage of a sensor that detects how polluted

is the air being sucked by the vacuum, this makes the robot to spend more time on

dirtier areas.

Probably the most successful practical application of home mobile robotics is the

“Roomba” vacumm cleaner by iRobot Co. (see Figure 1.1), in sales it surpases by far

its competitors. According to ofﬁcial statements it sold more than 2 million units by

2006. It is a small robot with a diameter of 34cm and 9cm high. Its main sensing

components are a contact bumper at the front and several infrared sensors located in

different places such as on the bottom to prevent it from falling down stairs or drops.

Its operation does not require a map of the area to be cleaned. It aims to cover

an area using simple algorithms such as spiral patterns, random bounce and wall fol-

lowing. The new generation of Roombas can sense dirtier areas and spend more time

there as well as calculate the time required to clean a room based on the amount of dirt

detected and the longest straight line they can perform without bumping into an object.

It has also the ability to recharge itself in a docking station.

In 2010, Neato Robotics started selling its vacuum cleaner that mechanically is

meant to be better with a more powerful vacuum system, also its simplicity allows the

user to press only one button to start operation and it has much more advanced features

than others on the market. Its main capability is to use a SLAM approach using a

specially designed low-cost laser range-ﬁnder to build a 2D map of its environment in

real-time. This permits path planning to clean a room in an optimum manner. Also it

charges itself and detects doors in order to clean a room before moving into another.

This robot is very affordable and can be used as a research platform as well.

Also in 2010, Evolution Robotics released to the market its sweep and mop robot

“Mint”. It has similar features to other robots of its kind, such as easy of usage and

1.2 Consumer Robotics: Current Status and Future Prospects 5

Roomba Neato V−R4000 Trilobite RC−3000

RC−3000 Karcher Yes IR, bump, dirt Preprogrammed motion

determined by sonars)

MODEL COMPANY SELF−CHARGING SENSORS CLEANING ALGORITHMS PRICE (GBP)

1200

dirt more accuracy in robot’s position)

Neato Neato Robotics Yes cliff, 2D laser, bump SLAM with path−planning

V−R4000 LG Yes cliff, bump, IR, gyro Preprogrammed motion (gyro gives 800

Roomba iRobot Yes cliff, bump, IR, dirt Preprogrammed motions 100

Trilobite Electrolux Yes cliff, bump, sonars Preprogrammed motion (Perimeter 900

350

Figure 1.1: Vacuum robots comparison

a small size. Its navigation system is based on the “NorthStar” technology. This is a

photonic localisation system that projects a unique IR light spot onto the ceiling that

is detected by a sensor in the robot and then triangulated to obtain an accurate position

and heading. As it goes, the robot builds a map of the environment allowing it to have

intelligent algorithms to cover the whole area in a minimum amount of time. One of

the problems of this technology is the distance between the detector and the light plane

which must not exceed 6m; so a network of projectors is needed to cover larger spaces.

The way the robot works is by simply attaching clean pads to the front-lower part of

the robot and by pressing start. In their ﬁrst version the difference between mop and

sweep modes was only a change in the robot’s motion (mopping is by going forward

and backwards). In the latest release “Mint plus” they add a liquid dispenser to fulﬁl

partially the needs of a mopping task.

There are other commercialised robots with rather different practical applications,

e.g:

• Mopping: “Scooba” by iRobot Co. mops the ﬂoor using similar operation modes

as “roomba”.

• Lawn mower: “Robomower” by Friendly robotics. It weights 22.5 Kg and it is

equipped with bumper sensors to ﬁnd obstacles, it also uses a wire that produces

a signal as a virtual wall to delimit the area. It has some drawbacks as: the area

has to be setup a priori, no obstacle avoidance is implemented and the obstacles

detected by the bumper sensor have to be rigid and with certain dimensions. A

more resent robot is “Robocut” by Brill, it was presented in 2006 claiming to

have great cutting mechanisms and an optimised guidance system, it is designed

Introduction

to operate in areas of ∼ 2500m

• Pool cleaner: “Polaris 145” is an automatic pool cleaner, which operates by

wondering around at the bottom of the pool cleaning it until it covers the whole

of the pool. Similar products are: “Aquabot” by Aqua Products and “Tiger

Shark” by Aquavac.

• Ironing Robot: “Dressman” by Siemens. It is designed to iron shirts by ﬁtting

them on its inﬂatable torso. The ironing process consists of ﬁlling the robot’s

torso with hot air.

The autonomy level of the robots now on the market is improving, but generally

still limited. There is a clear demand for products which can be trusted to perform

complex tasks without human training, set-up or monitoring.

Having specialised sensors for every task could help in obtaining more data about

the environment, but we would like to think also about the cost that the robot could

be built for and we think that there will be a continued move towards commodity,

standard sensing in robotics. In practical robotic applications we would like to increase

performance while also decreasing price; the success of the Roomba vacuum cleaner

compared to other products show the great difference in adoption that a cheap robot

can achieve. So attention should primarily be focused on algorithms to make better

real-time use of the information available from standard sensors.

None of the robots presented in the previous paragraphs has a camera as its main

outward-looking navigation sensor; or do not have cameras at all. These is due to

several reasons; but mainly because even though an image contains a vast amount of

information, analysing it and extracting all the data needed is not as straightforward

as obtaining depth measurements from a range sensor. Working with images requires

much processing before obtaining the needed data and also is subject to illumination

changes. Hence, for targeted applications such as a vacuum cleaner where only 2D in-

formation is needed to compute a map or to detect obstacles then using a range sensor

might be much easier. However, by using a camera in a robot, we can aim at much

more generality and a camera has almost limitless potential for more advanced capa-

bilities such as recognising humans or characterising a scene with richer information

that can be used in further tasks or applications. For instance, in a vacuum cleaner,

a camera could potentially be used as the only sensor to perform SLAM, detect steps

and doors, and depending on the colour and texture it might recognise the type of ﬂoor

or carpet and use a different mode more suitable to it.

1.2 Consumer Robotics: Current Status and Future Prospects 7

1.2.2 Mid-Term Potential

Over the next 5 years, we consider that the main developments in home robotics will

be to consolidate current research capabilities such as:

• Robots with robust autonomous navigation capabilities, able to autonomously

map complex real environments

• Robots with more advanced scene understanding: e.g. using wide angle vision

to quickly estimate the type or main dimensions of a room.

• Elements of object recognition coming into play: door recognition, target object

recognition, etc.

• Robot-computer interaction: the ability to interact with a person via gestures.

In this section some study cases will be analysed from the computer vision point

of view to extract important aspects that will help as a starting point for a computer

vision framework that could serve in future applications. We aim to conceive practical

applications that could have a serious time-saving impact on people’s lives. For the

purpose of this research, we assume the most general case where the environment

where the robot is moving in is not structured, and also that no beacons or landmarks

are added.

• Autonomous vacuum cleaner

It can be seen from the last section that the vacuum cleaners now on the mar-

ket are still limited in their capabilities, though with products like Trilobite and

Neato this is changing. What would happen if we could make a smart vacuum

cleaner that builds up a map automatically each time it works incrementing the

efﬁciency only vacuuming the areas where it has not been yet? This could min-

imise battery use and time. Depending on the accuracy of the map, the robot

could compare between a saved map and the actual map and give a report of

the areas that were not cleaned due to different obstacles. It could also have the

ability of interact with humans while it is working; a person could give it visual

signals such as: stop, clean again or pause. It could ﬁnd the best path to any

room or the charging station and charge itself. These overall capabilities require

a number of different modules to be developed and integrated, the main ones

being the following:

– SLAM

Localisation and navigation could be said to be the base for an autonomous

mobile robot. In vacuum cleaning, which requires precise and complete

Introduction

coverage of an area rather than just ‘A to B’ planning, accurate and robust

localisation is particularly critical. This module enables the robot to know

which parts of a room have been cleaned and which not yet, and the relative

locations of multiple rooms and its charging charging station.

Vacuum cleaning does require good accuracy and repeatability of localisa-

tion, though probably at the centimetre level rather than any more precise.

A SLAM solution which is at least locally of good metric quality is there-

fore suitable. The robot will be making repeated and loopy movement in a

restricted area, surrounded by furniture, so a wide clear of view for accu-

rate mapping and relocalisation will be important.

– Free space detection and obstacle avoidance

In a dynamic and complicated environment such as a house it is certainly

necessary to have obstacle avoidance routines on a robot. A vacuum

cleaner needs to identify all cleanable free space that needs to be cleaned

in a comprehensive manner. It should also be able to approach obstacles

very closely to clean along edges. A ﬁnal possible capability, though pos-

sibly very difﬁcult, would be actually to observe the ﬂoor itself to detect

automatically how clean it was.

– Human recognition and human interaction

Human interaction is possibly an important aspect in this application. The

robot could be able to identify people in the environment and through vi-

sual signals it could take actions such as stop, clean again or pause. An-

other important aspect which could add more relevance to our work is hav-

ing feedback from humans.

Different aspects have to be analysed before starting building the system. Prob-

lems such as illumination changes has to be taken in to account, illumination is a

difﬁcult problem when dealing with unstructured scenarios, the vacuum cleaner

looking into a shadow could thing that it is dirty or when it is under the sun light

from a window this could affect its perception of the surroundings, that is why

having a reliable system invariable to illumination changes would be really im-

portant, if that can not be achieved with high accuracy then artiﬁcial intelligence

algorithms could be applied in order to adapt itself to different times during the

day.

Another aspect to think about, is the necessity or not about depth information,

because the robot is only moving in 2D and different segmentation algorithms

and occupancy grids could be used to detect objects and represent the map then

I believe that for this task depth information could be not used.

剩余179页未读，继续阅读

Discoverhub

粉丝: 936
资源: 48

多相机计算机视觉下的机器人SLAM与导航技术

MULTICOL-SLAM - A MODULAR REAL-TIME MULTI-CAMERA SLAM SYSTEM.pdf

Multi-robot-SLAM-Code-master_slam_multirobotslam_多机器人_multirobot

论文研究-基于多传感器融合的移动机器人SLAM .pdf

ubuntu三维激光slam安装

ubantu18.04 slam安装

ros slam小车 自动导航 路径规划 代码

ROS官方文档关于navigation包的链接

UBUNTU ROS SLAM学习路径

智能消毒机器人的路径规划与消毒方式研究有相关的文献吗？

ROS2 humble怎么运行ros2 navigation stack

如何在建图后用rviz 进行距今规划

ROS机器人实现牛耕往复式路径规划的功能包

ROS的移动机器人地图构建与导航国内外研究现状

Ubuntu20.04思岚激光雷达导航

.运行 navigation 导航避障算法，结合雷达，超声波，深度摄像头的数据，在导航移动中避开相 应障碍物，到达指定目标点

ros 的有趣功能包实现

ros使用自己的导航包

4舵机运动模型 ros

最新资源

ros slam小车自动导航路径规划代码

.运行 navigation 导航避障算法，结合雷达，超声波，深度摄像头的数据，在导航移动中避开相应障碍物，到达指定目标点