MATLAB中的机器人视觉与控制：图像处理与边缘检测

5星 · 超过95%的资源需积分: 9 47 浏览量更新于2024-07-19 1 收藏 27.31MB PDF 举报

"Robotics, Vision and Control Fundamental Algorithm in MATLAB 2nd Edition Part 3" 本文主要探讨了在MATLAB环境中实现机器人技术、视觉处理及控制基础算法的第二版中的部分内容，特别是图像处理的两个关键算法：边界检测和hit-or-miss变换。 12.6.2 边界检测边界检测是图像处理中的一个重要环节，用于识别图像中物体的边缘。在文中，通过使用形态学操作——顶帽变换来实现这一功能。顶帽变换通过将原图像与经过结构元素腐蚀后的图像相减，来突出图像中的边界。例如，对于一个名为`clean`的图像（如图12.28c所示），我们使用圆形结构元素进行腐蚀操作： ```matlab eroded = imorph(clean, kcircle(1), 'min'); ``` 腐蚀操作使得图像中的每个物体边缘外侧的一像素被去除。接着，将腐蚀后的图像从原始图像中减去： ```matlab idisp(clean - eroded); ``` 这样就得到了一个围绕每个物体边缘的像素层，如图12.29所示。 12.6.3 hit-or-miss变换 hit-or-miss变换是形态学操作的一种变体，其结构元素包含零、一和don't care值。当结构元素中的零和一像素与图像像素完全匹配时，结果才为一。MATLAB工具箱提供了类似的函数来实现这个操作，如： ```matlab out = hitormiss(image, S); ``` 其中，`image`是输入图像，`S`是结构元素。hit-or-miss变换对结构元素中的每个像素进行检查，如果结构元素中的零或一与图像像素不匹配，结果将为零。如图12.30a、b、c所示，展示了匹配和不匹配的情况。这些算法在机器人视觉系统中有着广泛的应用，例如在目标检测、环境感知和导航等方面。通过MATLAB这样的强大工具，我们可以方便地实现和调试这些复杂的图像处理算法，为机器人系统的开发和优化提供便利。在实际应用中，结合机器人的传感器数据（如摄像头图像），这些算法可以用于识别和跟踪物体，确定它们的位置和形状，从而帮助机器人做出适当的决策。边界检测可以帮助提取关键特征，hit-or-miss变换则可用于特定模式的搜索和定位。随着MATLAB版本的更新，这些算法的效率和精度也在不断提升，进一步推动了机器人技术的发展。

414

Chapter 13 · Image Feature Extraction

415

13.1

Region Features

Image segmentation is the process of partitioning an image into application meaning-

ful regions as illustrated in Fig. 13.1. The aim is to segment or separate those pixels

that represent objects of interest from all other pixels in the scene. This is one of the

oldest approaches to scene understanding and while conceptually straightforward it

is very challenging in practice. A key requirement is robustness which is how grace-

fully the method degrades as the underlying assumptions are violated, for example

changing scene illumination or viewpoint.

Image segmentation is considered as three subproblems. The ﬁ rst is classiﬁ ca-

tion which is a decision process applied to each pixel that assigns the pixel to one of

C classes c ∈ {0  C − 1}. Commonly we use C = 2 which is known as binary classiﬁ -

cation or binarization and some examples are shown in Fig. 13.1a–c. The pixels have

been classiﬁ ed as object (c = 1) or not-object (c = 0) which are displayed as white

or black pixels respectively. The classiﬁ cation is always application speciﬁ c – for ex-

ample the object corresponds to pixels that are bright or yellow or red or moving.

Figure 13.1d is a multi-level classiﬁ cation where C = 28 and the pixel’s class is reﬂ ect-

ed in its displayed color.

The underlying assumption in the examples of Fig. 13.1 is that regions are homoge-

neous with respect to some characteristic such as brightness, color or texture. In prac-

tice we accept that this stage is imperfect and that pixels may be misclassiﬁ ed – sub-

sequent processing steps will have to deal with this.

The second step in the segmentation process is representation where adjacent pixels

of the same class are connected to form spatial sets S

… S

. The sets can be represent-

ed by assigning a set label to each pixel or by a list of pixel coordinates that deﬁ nes the

boundary of the connected set. In the third and ﬁ nal step, the sets S

are described in

terms of compact scalar or vector-valued features such as size, position, and shape.

13.1.1

Classification

The pixel class is represented by an integer c ∈ {0  C − 1} where C is the number

of classes. In this section we discuss the problem of assigning each pixel to a class.

In many of the examples we will use binary classiﬁ cation with just two classes corre-

sponding to not-object and object, or background and foreground.

13.1.1.1

Grey-Level Classification

A common approach to binary classiﬁ cation of pixels is the monadic operator

where the decision is based simply on the value of the pixel I. This approach is called

thresholding and t is referred to as the threshold.

Thresholding is very simple to implement. Consider the image

>> castle = iread('castle.png', 'double');

which is shown in Fig. 13.2a. The thresholded image

>> idisp(castle >= 0.7)

is shown in Fig. 13.2c. The pixels have been quite accurately classiﬁ ed as corresponding

to white paint or not. This classiﬁ cation is based on the seemingly reasonable assump-

tion that the white paint objects are brighter than everything else in the image.

Fig. 13.1.

Examples of pixel classiﬁ cation.

The left-hand column is the in-

put image and the right-hand

column is the classiﬁ cation. The

classiﬁ cation is application spe-

ciﬁ c and the pixels have been

classiﬁ ed as either object (white)

or not-object (black). The ob-

jects of interest are a the indi-

vidual letters on the sign; b the

yellow targets; c the red toma-

toes. d is a multi-level segmen-

tation where pixels have been

assigned to 28 classes that rep-

resent locally homogeneous

groups of pixels in the scene



13.1 · Region Features

418

Chapter 13 · Image Feature Extraction

and the result of applying this threshold is shown in Fig. 13.3c. The pixel classiﬁ cation

is poor and the highlight overlaps several of the characters. The result of using a higher

threshold of 0.75 is shown in Fig. 13.3d – the highlight is reduced, but not completely,

but some other characters are starting to break up.

Thresholding-based techniques are notoriously brittle – a slight change in illu-

mination of the scene means that the thresholds we chose would no longer be

appropriate. In most real scenes there is no simple mapping from pixel values

to particular objects – we cannot for example choose a threshold that would

select a motorbike or a duck. Distinguishing an object from the background re-

mains a hard computer vision problem.

One alternative is to choose a local rather than a global threshold. The Niblack

algorithm is widely used in optical character recognition systems and computes a

local threshold

where W is a region about the point (u, v) and

(·) and

(·) are the mean and standard

deviation respectively. The size of the window W is a critical parameter and should

be of a similar size to the objects we are looking for. For this example we make an

assumption about the scene, that the characters are approximately 50–70 pixels tall,

to choose a window half-width of 30 pixels

>> t = niblack(castle, -0.1, 30);

>> idisp(t)

where k =−0.1. The resulting local threshold t is shown in Fig. 13.4a. We apply the

threshold pixel-wise to the original image

>> idisp(castle >= t)

resulting in the classiﬁ cation shown in Fig. 13.4b. All the pixels belonging to the let-

ters have been correctly classiﬁ ed but compared to Fig. 13.3c there are many false

positives – nonobject pixels classiﬁ ed as objects. Later in this section we will discuss

techniques to eliminate these false positives. Note that the classiﬁ cation process is no

longer a function of just the input pixel, it is now a complex function of the pixel and

its neighbors. While we no longer need to choose t we now need to choose the param-

eters k and window size, and again this is usually a trial and error process that can be

made to work well for a particular type of scene.

Fig. 13.4. Niblack thresholding.

a The local threshold displayed as

an image; b the binary segmen-

tation

剩余284页未读，继续阅读

CSUFT0306

粉丝: 17
资源: 12

MATLAB中的机器人视觉与控制：图像处理与边缘检测

Robotics, Vision and Control (Fundamental Algorithms in MATLAB)

Robotics, Vision and Control (Fundamental Algorithms in Matlab)(2017最新版本，高清非扫描版)

Robotics, Vision and Control

robotics, vision and control fundamental algorithms in matlab csdn

robotics,vision and control-2nd edition

introduction to robotics mechanics and control 3rd edition

robotics, vision & control second edition

工业机器人matlab参考文献

robotics system toolbox 和robotics toolbox for MATLAB异同点

robotics system toolbox 和robotics toolbox for MATLAB具体描述二者的异同

最新资源