服务与现场机器人技术综述：Mejias, Corke & Roberts编

需积分: 10 38 浏览量更新于2024-07-18 收藏 68.9MB PDF 举报

《Field and Service Robot》是由Jonathan Roberts担任编辑的一本专著，它作为Springer Tracts in Advanced Robotics系列的第105卷，专注于实地和服务型机器人的研究与实践。这本书深入探讨了机器人技术的关键领域，包括运动学与动力学模型、控制算法以及运动规划策略。在本书中，作者Luis Mejias和Peter Corke共同编撰，集结了国际第九届会议的结果，这些会议反映了当前机器人技术的前沿进展。书中内容涵盖了机器人设计中的基础理论，如机械结构如何影响其运动性能，以及如何通过精确的运动学模型来理解和预测机器人的行为。运动学模型是理解机器人如何在空间中移动和操作的基础，它涉及到关节角度、位姿和速度的数学表示。动力学模型则进一步关注机器人在执行任务时的能量消耗、力的传递以及加速度分析。通过理解这些模型，工程师能够优化机器人的能量效率，确保其在复杂环境中能稳定且有效地工作。控制算法则是实现机器人自主决策和动作的关键，包括经典的PID控制器、模型预测控制（MPC）等，它们确保机器人能够准确响应外部输入并达到预设目标。运动规划算法部分，本书探讨了如何设计高效的路径规划和避障策略，使服务机器人能够在动态环境中避免碰撞，同时执行复杂的任务路径。这通常涉及到路径搜索算法（如A*算法）、轨迹优化以及实时决策制定，以适应不断变化的环境条件。此外，书后的编辑顾问团队由国际知名学者组成，他们的专业知识和经验为读者提供了更广阔的视野，展示了机器人技术在全球不同学术背景下的多元化发展。例如，来自斯坦福大学的Oussama Khatib，以其在人工智能领域的贡献，强调了机器学习和自主控制在服务机器人领域的潜力。《Field and Service Robot》是一本综合性的著作，不仅适合学生和研究人员，也对工业界工程师有着重要的参考价值，因为它提供了关于如何将理论知识转化为实际应用的实用指导。对于那些希望在机器人技术特别是实地服务应用领域深入探索的人来说，这本书是不可或缺的资源。

4 M.S. Bewley et al.

reduce the current dependence that marine scientists have on extrapolating

information from a relatively small number of sparsely labelled points.

1 Introduction

Autonomous Underwater Vehicles (AUVs) have made a signiﬁcant impact

on areas of marine science that require an understanding of the sea ﬂoor [2,

3, 4]. Common practice is shifting from using human divers with hand-held

cameras, to sending AUVs equipped with stereo cameras and other sensors

to capture benthic images. These AUVs are capable of capturing far more

data, both in the type (various sensor modalities) and volume (a several

hour mission can collect tens of thousands of spatially registered stereo image

pairs). In addition, the imagery is geo-referenced far more precisely, and data

can be gathered from beyond diver depths. While using an AUV improves

the raw data in many ways, the sheer volume of it introduces a new problem

for interpretation.

For marine scientists studying the location, distribution and coverage of

benthic organisms and morphology, the state of the art is to take a small

subset of images from an AUV survey, and manually label the content. For

the data set used in this paper [5], biological species and physical formations

under 50 randomly selected pixels from every 100

image were labelled.

The scientists then extrapolate from this subset, to make inferences about

the ecosystems and populations in the geographic area. Figure 5 shows an

example image from our data set (described in Section 3) where 50 randomly

selected pixels have been labelled using the CPCe labelling software [6].

Fig. 1 Example AUV image, with CPC point labels

Hierarchical Classiﬁcation in AUV Imagery 5

The ultimate goal of research in this ﬁeld is to automate the labelling

process, such that the vehicle returns a semantic map of the environment,

rather than simply providing the raw visual and sensory data. From a robotics

perspective, this requires the robotic system to take on the translation and

communication eﬀort with which the human users are currently burdened.

Recent work [1] describes a solution to a simpliﬁed problem of automatically

classifying a single species (Ecklonia Radiata commonly known as “Kelp”),

based on supervised learning algorithms applied to a large AUV dataset. In

this paper, we extend the problem to multiple species and physical morpholo-

gies, using the same data set.

2 Hierarchical Classiﬁcation

In order to perform classiﬁcation on a large set of benthic classes, a diﬀerent

approach is required. Supervised machine learning is typically performed on

binary problems (such as presence/absence detection) [7, 3], or multi-class

problems where there are a set of strict alternatives [8, 9]. In this study, we

examine solutions to the more complex problem of predicting a hierarchy of

classes, which is the semantic output that is desired by the marine science

community that uses this data set (see Section 3 for details). The contribution

of this paper is to demonstrate and evaluate the techniques used in other

areas (computer vision, text classiﬁcation and medical diagnosis) applied to

hierarchical classiﬁcation in benthic images.

2.1 Classiﬁcation

In a recent survey, Silla Jr and Freitas [10] extensively reviewed the problem

of supervised hierarchical classiﬁcation and deﬁned a range of methods of

dealing with the hierarchy of the data. They deﬁned three main approaches:

1. Ignore the hierarchy and perform ﬂat classiﬁcation (Figure 2a)

2. Use a network of several local binary classiﬁers for various regions of the

hierarchy (Figure 2b)

3. Use a single classiﬁer but encode the hierarchical structure somehow in

the data (Figure 2c).

From the list of methodologies presented, we employ the Local Classiﬁer per

Node approach of Figure 2b. In this approach, which is by far the most

used in the literature [10], each node in the classiﬁcation tree has a binary

classiﬁer that is trained to distinguish that class from others. This approach

is relatively straightforward to implement, but has several advantages:

• It is possible to either output the single node in the hierarchy which rep-

resents the best prediction, or to query the probability of a given image

point being a member of any given node.

Hierarchical Classiﬁcation in AUV Imagery 7

siblings of the ancestor nodes). Using the example from before, the negative

training samples will come from nodes 2.2, 2.2.1 and 2.2.2 only.

The expected performance diﬀerence between these two policies is not

obvious, with no consistent winner found in [11]. On one hand, the inclusive

policy ensures that each node is as informed as possible, and should be able

to deal better with classifying instances that belong elsewhere in the tree

(as it has seen examples of all the data). On the other hand, the siblings

policy solves a much more speciﬁc problem (distinguishing a node’s class

from its siblings), and may give better discriminative performance between

these classes. Such nodes will, however, be less informed about instances that

belong elsewhere in the tree. An inherent advantage of the siblings approach

is that far less training data is required for nodes deeper in the tree, which

becomes signiﬁcant when the tree is large.

2.3 Prediction

After training the network of classiﬁers, the decision remains of how to predict

the class of a given image point. As we require complete consistency in the

hierarchical labels (such that there is a single, unbroken chain of classiﬁers

predicting positive results from the root to the deepest node), the simplest

choice described in [10] is chosen, which we refer to as maximum probability

switching (MPS). An instance starts at the root node, and ﬂows to the leaf

node with the highest prediction probability (akin to performing one-vs-rest

classiﬁcation at each node). This technique forces prediction down to the

leaf node level. We can remove this constraint by stopping a prediction from

moving further down the tree when the maximum predicted probability falls

below some threshold (say 0.5 for the typical cut-oﬀ of a binary classiﬁer).

We also test an alternative approach: the use of a simple probabilistic graphi-

cal model (PGM). Here the class tree also represents the independence relations

in the PGM, and we assume the conditional probability of node membership

is given by the probabilistic predictions of the classiﬁers

. This allows exact

Bayesian inference to be performed trivially, by multiplying probabilities of a

node’s ancestors to obtain the net probability of membership.

2.4 Performance Measure

Lastly, a robust performance metric is needed; ideally a single number to

evaluate the performance on an entire tree. The closest performance measure

commonly used in the literature is the hierarchical f1-score [12]. Each instance

has multiple counts of true/false positives/negatives, as each node in the chain

This assumption is at least reasonable for the siblings training policy, as the train-

ing sets only include those instances which we know belong to the parent node. The

output of the classiﬁers therefore represents e.g. “the probability that this instance

is Algae, given we know it is an instance of Biota”.

8 M.S. Bewley et al.

(a) True predic-

tion: (TP, TP,

TP)

(b) False predic-

tion: (TP, TP,

FP, FN)

diction due to

thresholding:

(TP, TP, FN)

(d) Modiﬁed

prediction: (TP,

TP)

Fig. 3 The hierarchical f1-score performance metric. The connected circles repre-

sent classiﬁer nodes, with the root on the left, and leaf on the right. Grey circles are

ground truth, white circles are prediction. Figure 3a shows the case of the classiﬁer

network correctly predicting a node of depth 3 in the hierarchy. Figure 3b repre-

sents the ﬁrst two levels being predicted correctly, whereas the classiﬁer network

incorrectly chooses a sibling of the true class at depth 3.

of true class nodes is compared to the chain of predicted nodes. Figures 3a

and 3b illustrate two scenarios. In the former case, the prediction down the

entire tree is accurate. In the latter case the last prediction is incorrect. This

results in a False Negative (FN) detection as the correct node was missed as

well as a False Positive (FP) detection for an incorrect node being detected.

When employing thresholded MPS it is possible that the prediction stops

earlier up in the tree than the ground truth. This occurs when the child

classiﬁer prediction probability is below the set threshold. In such an instance,

the missed nodes are calculated as FNs. This is illustrated in Figure 3c.

Finally, there can be cases where the predicted class is more speciﬁc (lower

down the tree) than the ground truth class. This can occur when the ground

truth label is not at the leaf level. In such an instance the hierarchical f1-

score commonly used in the literature would penalise the prediction with

a FP. However, there are instances where this is not fair; e.g. the classiﬁer

predicts “Labrador” while the label says “Dog”. “Labrador ” is of course a

breed of “Dog”, and may be a more accurate description. This is true for

our labels of the underwater species, where the labeller does not always label

the instance to the leaf-node level. This can occur due to poor image quality,

inability of the labeller to recognise the species or cost and eﬀort constraints

if more detailed labels are not required for the given research project.

We therefore modify the metric such that if the predicted class is more

speciﬁc than the ground truth (Figure 3d), we do not reward or penalise any

results deeper than the deepest known class.

2.5 Descriptors

A subtlety of the Local Classiﬁer per Node approach is in the selection of image

features. In ﬂat multi-class classiﬁcation (Figure 2a) and global classiﬁcation

剩余522页未读，继续阅读

凌寒画家

粉丝: 14
资源: 13

服务与现场机器人技术综述：Mejias, Corke & Roberts编

simple-robot/simpler-robot

MATLAB Code of Artificial Potencial Field Method for Robot Path Planning

The ZigBee Based Wireless Sensor and Actor Network in Intelligent Space Oriented to Home Service Robot

cp -av /media/robot/_home/* ./ 这是什么意思

The specified source space "/home/robot/catkin_ws/src" does not exist

接口为：https://oapi.dingtalk.com/robot/create 实例

fatal: repository 'https://github.com/lg609/aubo_robot/blob/master/aubo_description/urdf/aubo_i5.git/' not found

Linking CXX executable /home/yu/my_ros_can_car/devel/lib/can_ros_robot/can_ros_robot

最新资源