机器学习与数据分析：最新研究与应用

需积分: 9 79 浏览量更新于2024-07-18 1 收藏 6.38MB PDF 举报

"《机器学习与数据分析》是电气工程领域讲座笔记系列的第48卷，汇集了在会议中参与的杰出研究人员撰写的16篇修订和扩展的研究文章。本书涵盖了专家系统、智能决策、基于知识的系统、知识提取、数据分析工具、计算生物学、优化算法、实验设计、复杂系统识别、计算建模以及工业应用等多个主题，提供了机器学习和数据分析领域的最新进展，并且可以作为该领域研究者和研究生的重要参考教材。" 在《机器学习与数据分析》这本书中，作者们深入探讨了一系列关键的IT知识点： 1. **机器学习 (Machine Learning)**：这是人工智能的一个分支，通过让计算机在数据的帮助下自我学习和改进，而无需明确编程。书中的内容可能包括监督学习、无监督学习、半监督学习、强化学习等方法，以及各种算法如决策树、神经网络、支持向量机和随机森林等。 2. **数据分析工具 (Data Analysis Tools)**：书中可能涵盖了用于数据预处理、探索性数据分析、模型训练和验证的各种工具和技术，如Python的Pandas和NumPy库、R语言、SQL数据库查询、数据可视化工具（如Matplotlib和Seaborn）以及大数据分析平台（如Hadoop和Spark）。 3. **知识提取 (Knowledge Extraction)**：这是一个从大量数据中抽取有意义信息的过程，包括文本挖掘、模式识别、特征选择和概念建模。这部分内容可能会讨论如何从非结构化数据中抽取知识并转化为可用于决策的结构化形式。 4. **计算生物学 (Computational Biology)**：结合机器学习和数据分析，该领域关注生物系统的数学建模和计算模拟。可能涉及基因组学、蛋白质组学、药物发现和疾病预测等领域。 5. **优化算法 (Optimization Algorithms)**：在机器学习中，优化算法用于调整模型参数以最小化或最大化特定目标函数。可能涵盖梯度下降、牛顿法、遗传算法、粒子群优化等。 6. **实验设计 (Experiment Designs)**：在机器学习和数据分析中，有效的实验设计能帮助我们更好地理解数据和模型性能。这可能包括交叉验证、A/B测试、超参数调优等策略。 7. **复杂系统识别 (Complex System Identification)**：这一部分可能涉及到如何理解和建模复杂的动态系统，如神经网络、生态系统或金融市场。 8. **计算建模 (Computational Modeling)**：利用计算机模拟现实世界现象，可能是通过建立数学模型并应用机器学习技术进行预测和解释。这本书不仅提供了理论基础，还强调了实际应用，适合对机器学习和数据分析感兴趣的科研人员和学生作为深入研究和实践的指南。通过对这些主题的深入学习，读者将能够掌握解决实际问题的技能，例如预测、分类、聚类和异常检测等。

8 S.E. Ghobadi et al.

100

200

300

400

500

600

40 50

Fig. 1.4 Segmentation in 2D/3D images. Top Left: low resolution range image from TOF sensor.

Top Middle: high resolution 2D image. Top Right: modulation amplitude image. Bottom Left:3D

rescaled segmented image. Bottom Middle: rescaled segmented image using fusion of range and

modulation amplitude data. Bottom Right: 2D segmented image result from mapping

In this section, we describe tracking with classiﬁer more in detail by applying

a supervised classiﬁer based on AdaBoost to 2D/3D videos in order to detect and

track the desired object. After segmentation of the image which was described in the

previous section, in the next step the Haar-Like features are extracted and used as

the input data for the AdaBoost classiﬁer. Haar-like features which have been used

successfully in face tracking and classiﬁcation problems encode some information

about the object to be detected. For a much more in depth understanding the reader

is referred to [11].

However, there are two main issues in real time object detection based on Haar-

Like features and using AdaBoost technique. The ﬁrst issue is that background noise

in the training images degrades detection accuracy signiﬁcantly, esp. when it is a

cluttered background with varying lighting condition which is the case in many real

world problems. The second issue is that computation of all sub-windows (search

windows) in an image for every scale is too costly if the real time constraints are to

be met. The fundamental idea of our algorithm is to address the solution to these

problems using fusion of 3D range data with 2D images. In order to extinguish the

background issue from object recognition problem, the procedure of object detec-

tion is divided into two levels. In the low level we use range data in order to: (i)

Deﬁne a 3D volume where the object of interest is appearing (Volume of Interest)

and eliminate the background to achieve robustness against cluttered backgrounds

and (ii) Segment the foreground image into different clusters. In the high level we

map the 3D segmented image to its corresponding 2D color image and apply Viola-

Jones method [11] (searching with Haar-Like features) to ﬁnd the desired object in

the image. Figure 1.5 shows some examples of this procedure for hand detection

and tracking.

10 S.E. Ghobadi et al.

1.5 Real Time Hand Based Robot Control Using 2D/3D Images

Nowadays, robots are used in the different domains ranging from search and res-

cue in the dangerous environments to the interactive entertainments. The more the

robots are employed in our daily life, the more a natural communication with the

robot is required. Current communication devices, like keyboard, mouse, joystick

and electronic pen are not intuitive and natural enough. On the other hand, hand

gesture, as a natural interface means, has been attracting so much attention for

interactive communication with robots in the recent years [2–4, 9]. In this con-

text, vision based hand detection and tracking techniques are used to provide an

efﬁcient real time interface with the robot. However, the problem of visual hand

recognition and tracking is quite challenging. Many early approaches used position

markers or colored gloves to make the problem of hand recognition easier, but due

to their inconvenience, they can not be considered as a natural interface for the

robot control. Thanks to the latest advances in the computer vision ﬁeld, the recent

vision based approaches do not need any extra hardware except a camera. These

techniques can be categorized as: model based and appearance based methods [7].

While model based techniques can recognize the hand motion and its shape exactly,

they are computationally expensive and therefore they are infeasible for a real time

control application. The appearance based techniques on the other hand are faster

but they still deal with some issues such as: complex nature of the hand with more

than 20 DOF, cluttered and variant background, variation in lighting conditions and

real time computational demand. In this section we present the results of our work

in a real time hand based tracking system as an innovative natural commanding

system for a Human Robot Interaction (HRI).

1.5.1 Set-Up

Set-up mainly consists of three parts: (1) A six axis, harmonic driven robot from

Kuka of type KR 3 with attached magnetic grabber. The robot itself has been

mounted onto an aluminium rack along with the second system component. (2) A

dedicated robot control unit, responsible for robot operation and communication by

running proprietary software from Kuka



company. (3) The main PC responsible

for data acquisition from 2D/3D imaging system (MultiCam) and running the al-

gorithms. Communication between the robot control unit and the application PC is

done by exchanging XML-wrapped messages via TCP/IP. The network architecture

follows a strict client server model, with the control unit as the client connecting to

the main PC, running a server thread, during startup.

12 S.E. Ghobadi et al.

Table 1.1 Confusion table

for hand detection system

Hand Non-hand

Hand 2,633 87

Non-hand 224 2,630

Sum 2,857 2,717

1.5.3 Experimental Results

For the Hannover Fair 2008, a simple task had been deﬁned to be performed by the

visitors and to put the system’s performance under the test as follows: Commanding

the robot to move in six directions using moving the hand with any kind of posture

in the corresponding directions, picking up a metal object with the magnet grabber

using palm to ﬁst gesture, moving the object using the motion of the hand and ﬁnally

dropping it in the deﬁned areas with palm to ﬁst gesture. It turned out that the system

handling has been quite intuitive, since different people have been able to operate

the robot instantly. In terms of reliability the whole system worked ﬂawlessly during

the complete time exposed at the fair. For training of the classiﬁer we took 1037

positive hand images from seven people, and 1,269 negative images from non-hand

objects in our lab environment. Using OpenCV we trained our classiﬁer with 20

stages and window size of 32  32. Although the classiﬁer was trained under the

lab conditions, it worked quite well under the extreme lighting conditions at the fair.

In order to analyze the performance of the system, we recorded the results of hand

detection from our GUI in the video format while different users were commanding

the robot. Likewise, we moved the camera and took the videos from the environ-

ment. These videos are labeled as “Positive” and “Negative” data. While positive

stands for the hand, the negative represents the non-hand objects. The data were ac-

quired using a PC with dual core 2.4 GHz CPU. The exposure time for 3D sensor

was set at 2ms while for 2D sensor it was about 10 ms. Under these conditions,

we had about 15 detected images (including all algorithms computational time) per

second. The confusion matrix derived from these videos with 2857 hand images and

2717 non-hand images is shown in Table 1.1. As it can be calculated from this table,

the system has a Hit Rate of 0.921, False Positive Rate of 0.032 and the recognition

accuracy of 94.4%.

1.6 Conclusion

In this work we study some aspects of object detection and tracking using 2D/3D

Images. These images are provided by a monocular 2D/3D vision system, so-called

MultiCam. The principle of this camera system as well as the registration and fusion

of 2D/3D data are discussed. This work is concluded with some results of a real time

hand based robot control application which was demonstrated at Hannover fair in

Germany in 2008.

剩余240页未读，继续阅读

cjzhu_mm

粉丝: 6
资源: 12

机器学习与数据分析：最新研究与应用

Data Analysis, Machine Learning and Applications

Machine Learning and Security: Protecting Systems with Data and Algorithms

Machine Learning Approaches ot Facial and Text Analysis

Machine Learning for Audio, Image and Video Analysis 2nd

Machine Learning for Multimedia Content Analysis

Data_Analysis_And_Machine_Learning_Projects

A Machine Learning Analysis of COVID-19 Mental Health Data_COVID

Machine Learning and Deep Learning with Python, scikit-learn, and TensorFlow

Big Data Analysis and Deep Learning Applications

MATLAB Path and Machine Learning: Optimizing Path Settings to Enhance Machine Learning Code ...

最新资源