A Multiple Feature-Based Image-Switching Strategy in Visual Sensor Networks

Abstract

Generally, one fixed camera is used to take still or dynamic images and extract proper information from the captured images. However, the process of analyzing images through the use of one camera is very sensitive to neighboring environmental factors, such as illumination, background, and noise; thus, it is hard to guarantee precision. To extract proper information from images more precisely in visual sensor networks, this paper proposes an image-switching strategy where, among different types of installed cameras, the one camera best suited to neighboring circumstances is chosen. The proposed strategy is to first receive initial images as input data and then extract multiple features representing neighboring circumstances from the input images. Subsequently, it is to define the neighboring circumstances metric, which is the weighted sum of the extracted features, and to dynamically switch cameras to obtain images in accordance with the neighboring circumstances. The results of the experiment show that the proposed dynamic switching strategy reliably chooses, from among different cameras, the one camera that is best suited to the neighboring circumstances.

1. Introduction

With the rapid development of visual sensors, wireless communication, pattern recognition, and relevant technologies, visual sensor networks, which are widely deployed, have been distributed highly efficiently in terms of cost and practicability [1–3]. The visual sensor networks are usefully applicable to a wide range of commercial and military application programs, such as applications for video monitoring, smart home, traffic monitoring, movement tracking, and terror prevention [4–7].

A visual sensor network can contain various factors, including different types of sensors, a dedicated communication network, and a proper routing technique for power supply. Among them, a camera for taking scenes is very important because it greatly influences the performance of its accompanying image-processing work. Generally, the cameras used most in computer vision and pattern recognition fields are stereo cameras [8], structured-light cameras [9], TOF (time-of-flight) cameras [10], infrared cameras [11], and thermal imaging cameras [12].

The stereo camera [8] uses two lenses to discern an object three-dimensionally, just as human eyes do. The stereo camera is one of the most widely used camera types but is unable to discern an object when illumination is too bright or too dark. Additionally, the camera is unable to determine corresponding points for objects, such as a wall without texture, and thereby fails to obtain distance information. The structured-light camera [9] can be interpreted in the same way of geometrically replacing one lens of a stereo camera with a projector. The structured-light camera is able to determine corresponding points, even in conditions of poor illumination and texture, but has the disadvantage of slow speed. The TOF camera [10] radiates light-wave energy, such as ultrasound and laser, to an object and calculates the returning time of the light wave from the object to measure a distance from it. Therefore, once circumstances change, it is hard for such a camera to analyze a sound wave. The infrared camera [11] uses infrared rays to create images in a manner similar to general cameras, which use visible rays to make images. In other words, the infrared camera detects infrared energy radiated out of an object by using its detector and then turns the detected energy into an image. The infrared camera is able to obtain the same quality of images at night as in the day and to obtain images at a long distance. However, such a camera is expensive. The thermal imaging camera [12] is a device that visually displays thermal energy radiated out of a subject's surface. The camera is applied to various areas, but the higher resolution the camera has, the more it costs.

Most conventional image-processing techniques use one type of camera to capture scenes, making it difficult to obtain images optimized for neighboring circumstances, such as illumination, background, and noise. Therefore, this paper discusses the installation of multiple cameras in a visual sensor network environment and proposes a strategy to dynamically choose the most suitable camera from multiple cameras on the basis of multiple features. Figure 1 illustrates the overview of the dynamic image-switching algorithm proposed in this paper.

Figure 1

Overview of the proposed system.

As shown in Figure 1, the proposed dynamic switching strategy first receives initial images as input data and extracts various features representing neighboring circumstances from the input images. The strategy then defines the metric in the form of the weighted sum of the extracted features and dynamically switches cameras in accordance with neighboring circumstances. In other words, the strategy is to reliably choose the camera that is best suited to neighboring circumstances among the installed multiple cameras, as well as to capture and analyze scenes with the chosen camera.

The rest of the paper is organized as follows. Section 2 describes previous studies related to the cameras used in the computer vision area. Section 3 explains the technique to extract multiple features representing neighboring circumstances. Section 4 presents the algorithm to switch multiple cameras dynamically. Section 5 shows the results of the experiment conducted to compare and evaluate the performance of the proposed strategy. Section 6 presents the conclusion of this paper and proposes future research directions.

2. Related Works

With the development of multimedia technologies and the falling cost of relevant hardware devices, various types of cameras have been released and applied to many different areas. Of the cameras currently being used, the following are those most widely used.

The stereo camera is one of the most generally used camera types [8]. Such a camera has a separate image sensor and a film frame for each lens, consisting of two or more lenses. Accordingly, the stereo camera employs a time differential between two lenses to obtain three-dimensional distance information, just as human eyes do. Stereo cameras may be used for making stereo views and 3D pictures for movies or for range imaging. Usually, the distance between two lenses in a stereo camera is set to the distance between two human eyes. The baseline of a stereo camera may be longer than the distance, but there is no guarantee of naturalness of the images. Although stereo cameras are universally used, they are unable to determine corresponding points for objects such as a wall without texture. Therefore, the camera fails to obtain depth information [13].

In the case of a structured-light camera, as shown in the system architecture of Figure 2, one of the stereo vision system's two camera lenses is replaced with a device to project light. By projecting the geometrically patterned light, known as structured light, the camera provides corresponding points of stereo vision to easily solve the problem of corresponding points [9]. Such a structured-light-based distance measurement technique has smaller operations to solve the issue of corresponding points and provides multiple corresponding points; thus, it is capable of calculating a large amount of distance data. However, compared to other techniques, it takes more time [14].

Figure 2

Principle of structured-light camera.

The basic principle of a TOF camera is to radiate a strong ray of light forward from the camera and detect its returning reflected light to measure a distance [10]. For the radiating energy of a TOF camera, ultrasound waves are used as sound waves and laser lights are used as light waves. In the case of a sensor using ultrasound waves, it is hard to measure a distance at one accurate point in the real world because of the characteristics of ultrasound waves, and only a representative distance value within a given range of angles in the real world is calculated. In the case of the time delay method using a laser, such a sensor is a one-dimensional sensor type to obtain distance information at one point of the real world through one measurement. In this case, multiple measurements are required to expand two-dimensionally or three-dimensionally [15].

The infrared camera detects infrared energy radiated from an object by using its detector and turns the detected infrared energy into electrical signals to display images two-dimensionally [11], as shown in Figure 3. Infrared light is not invisible and is an electromagnetic wave, which is longer than that of red color light. The infrared camera is made by combining a lens with a sensor that responds to infrared rays. Generally, a CCD or CMOS sensor responds to both visible and infrared rays. Therefore, by applying a filter to a black and white camera, it is possible to pick out the infrared rays; with the control of a gain, invisible infrared regions are presented. Most infrared cameras support unicolor implementation, but some support multicolors. Generally, the infrared camera is capable of obtaining the same quality images at night as in the day and can obtain images at a long distance, but it is rather expensive [16].

Figure 3

Principle of infrared camera.

The thermal imaging camera uses infrared lights, which are radiated out of an object's surface, to look at the surface state through temperature distribution imaging [12]. This process is made possible because thermal energy changes infrared waves or speed. Usually, the infrared light emitted by an object with heat has different waves than the infrared light of a cold object. An infrared sensor converts such a wave difference into an image. Generally, the higher heat an object has, the more infrared light it emits. The brightest color of a thermal imaging camera is traditionally white, its intermediate temperature is generally red or yellow, and its darkest color is blue; temperature-related scales are displayed next to an image. The thermal imaging camera features fast and visible diagnosis outputs. However, thanks to expensive infrared sensors and optical lenses, the thermal imaging camera's production cost is relatively high and the camera can be damaged by the temperature of the object being measured [17].

3. Extraction of Multiple Features

Usually, the work of extracting main features robustly from images is very important in subsequent image processing [18]. To find the features of the circumstances around a camera, this paper extracts three features: illumination, edge, and depth. In addition, to minimize the influence of noise at the time of extracting the features, an input image is first split into regular square blocks with $n \times n$ pixels, and then the features are extracted from each unit block [19]. Additionally, the three features are normalized, with values from 0 to 1.

For the illumination feature, this paper uses the I value in the $Y I Q$ color space. Generally, the $R G B$ color space is used for color image processing. However, for image processing, the $R G B$ color model needs all three (R, G, and B) colors to be processed, thus making it inefficient. Moreover, the $R G B$ color space is not only sensitive to bright changes in an image but can have different color values depending on the size and direction of illumination [20]. For these reasons, this paper employs the $Y I Q$ color space, which is known to be relatively robust with regard to illumination changes. The $Y I Q$ color space is induced from the $Y U V$ color space. Y (Luminance) represents the intensity of luminance, and I (Inphase) and Q (Quadrature) represent color and chroma in combination with the two factors. Equation (1) is used to transform the $R G B$ color space into the $Y I Q$ color space [21]:

\begin{matrix} [\begin{bmatrix} Y \\ I \\ Q \end{bmatrix}] = [\begin{bmatrix} 0.30 & 0.59 & 0.11 \\ 0.60 & - 0.27 & - 0.32 \\ 0.21 & - 0.52 & 0.31 \end{bmatrix}] [\begin{bmatrix} R \\ G \\ B \end{bmatrix}] . \end{matrix}

(1)

In this paper, the illumination feature of an image captured at a time is defined as in (2). Here, N represents the total number of blocks in the image; $Y^{i} (x, y)$ is the Y value at the x and y coordinates of the ith block; $Y_{μ}^{i}$ is the mean of the Y values of the ith block:

\begin{matrix} E_{i n t e n s i t y} = \frac{1}{N} \times \sum_{i = 1}^{N} Y_{μ}^{i}, \\ Y_{μ}^{i} = \frac{1}{n^{2}} \times \sum_{x = 1}^{n} \sum_{y = 1}^{n} \frac{Y^{i} (x, y)}{255} . \end{matrix}

(2)

For the edge feature, the Canny edge [22] is extracted from an input image. Edge means a pixel located at the border of two areas where luminance intensity sharply changes. Generally, when the edge extracted from an image is used, it is possible to find the location of an object in the image and obtain information on the shape and size of the extracted object. In particular, it is possible to identify how much texture exists in the image. To obtain the Canny edge, this paper first convolves a Gaussian kernel—as shown in Figure 4—in the input image $I (x, y)$ for smoothing.

Figure 4

Gaussian function.

Next, a differential operator is applied to the smoothed image to calculate the sizes of the vertical and horizontal edge; thus, the size and direction of the edge are calculated. Lastly, a non-maximum suppression process (to detect the edge with pixel thickness) and a double thresholding process (to connect edge points) are performed [23]. In this paper, the edge feature of a given image is defined as shown in (3). Here, $C^{i} (x, y)$ represents the size of the Canny edge at the x and y coordinates of the ith block; $C_{μ}^{i}$ is the mean size of the Canny edges in the ith block; $C_{m a x}$ is the maximum value of the Canny edge:

\begin{matrix} E_{e d g e} = \frac{1}{N} \times \sum_{i = 1}^{N} C_{μ}^{i}, \\ C_{μ}^{i} = \frac{1}{n^{2}} \times \sum_{x = 1}^{n} \sum_{y = 1}^{n} \frac{|C^{i} (x, y)|}{C_{m a x}} . \end{matrix}

(3)

This paper uses the depth feature [24] as the third feature. The basic concept of extracting depth information from left and right input images is illustrated in Figure 5.

Figure 5

Overview of depth extraction.

As shown in Figure 5, let us assume that P is one point in the real world; $x_{l}$ and $x_{r}$ are the coordinates in which P is placed in the left and right images; f is the focal length of the camera; T is the baseline of the camera; Z is the depth information to extract. In this case, the triangles $(p_{l}, P, p_{r})$ and $(O_{l}, P, O_{r})$ are similar to each other such that (4) is established. When (4) is rearranged to make Z the subject, it is written as in (5), and thus, the depth information can be extracted. In (5), $d = x_{r} - x_{l}$ represents disparity information [25]:

\begin{matrix} \frac{T + x_{l} - x_{r}}{Z - f} = \frac{T}{Z}, \end{matrix}

(4)

\begin{matrix} Z = f \times \frac{T}{d} . \end{matrix}

(5)

In this paper, the depth feature of a given image is defined as in (6). Here, $Z^{i} (x, y)$ represents the depth value at the x and y coordinates of the ith block of the image; $Z_{μ}^{i}$ is the mean of the depth values of the ith block; $Z_{m a x}$ represents the maximum value of depth:

\begin{matrix} E_{d e p t h} = \frac{1}{N} \times \sum_{i = 1}^{N} Z_{μ}^{i}, \\ Z_{μ}^{i} = \frac{1}{n^{2}} \times \sum_{x = 1}^{n} \sum_{y = 1}^{n} \frac{Z^{i} (x, y)}{Z_{m a x}} . \end{matrix}

(6)

4. Weight-Based Image-Switching Strategy

To combine the three features extracted in the previous step, this paper uses weighting factors [26, 27] to define the complexity metric of neighboring circumstances in the sum of their weights. Subsequently, the defined metric is used to dynamically choose, from among multiple cameras, the one camera best suited to the neighboring circumstances. In this paper, the complexity metric is extracted from the initial input image with the use of

\begin{array}{l} E (α, β, γ) = α \times (1 - |E_{i n t e n s i t y} - \frac{1}{2}| \times 2) + β \times E_{e d g e} \\ + γ \times (1 - |E_{d i s t a n c e} - \frac{1}{2}| \times 2), \end{array}

(7)

where

0 \leq α, β, γ \leq 1

α + β + γ = 1

In (7), α, β, and γ are the weighting factors used to determine the importance of each term. Each weight is a value between 0 and 1, and the sum of the three weights is 1. In (7), as shown in Figure 6(a), if $E_{i n t e n s i t y}$ equals 0 or 1, then the term of illumination feature of $E (α, β, γ)$ becomes a minimal value; if $E_{i n t e n s i t y}$ equals 0.5, then the term of illumination feature becomes a maximum value. In regard to the edge feature, as shown in Figure 6(b), if $E_{e d g e}$ is 0, the term of edge feature of $E (α, β, γ)$ becomes a minimal value; if $E_{e d g e}$ is 1, then the term of edge feature is a maximum value. Regarding the distance feature, as shown in Figure 6(c), if $E_{d i s t a n c e}$ is 0 or 1, then the term of distance feature of $E (α, β, γ)$ becomes a minimal value; if $E_{d i s t a n c e}$ is 0.5, the term of distance feature is a maximum value.

Figure 6

Changes in terms for each feature.

After the complexity metric of circumstances is extracted with the use of (7), the following if-else rule based on two thresholds is applied to dynamically switch cameras. In other words, when the complexity metric $E (α, β, γ)$ is more than the threshold TH2, the stereo camera is activated; when $E (α, β, γ)$ is between the thresholds TH1 and TH2, the TOF camera is activated; when $E (α, β, γ)$ is less than TH1, the infrared camera is activated. The two thresholds used in the following rule play a role in splitting a range of the complexity metric into three sections, which are empirically determined by the repetitive experiments of this paper (see Algorithm 1).

Algorithm 1

IF $(TH 2 \leq E (α, β, γ))$

THEN stereo camera is activated

ELSE IF $(TH 1 \leq E (α, β, γ) ≺ TH 2)$

THEN TOF camera is activated

ELSE IF $(E (α, β, γ) ≺ TH 1)$

THEN infrared camera is activated

END

The dynamic camera-switching technique, based on the defined rule with the complexity metric and thresholds, can be described similarly to the method of choosing a camera based on how large the extracted three feature values are, as shown in Table 1.

Table 1

Choice of camera according to features.

Camera	Feature
Camera	Illumination	Edge	Distance
Stereo camera	Neither bright nor dark	Plentiful	Neither near nor distant
TOF camera	Bright or dark	Poor	Near or distant
Infrared camera	Very bright or very dark	Very poor	Very near or very distant

Figure 7 illustrates the flowchart of the proposed algorithm to dynamically select a camera on the basis of the complexity metric of circumstances.

Figure 7

Dynamic switching process of images.

5. Experimental Results

For the experiment of this study, a computer that has a Pentium Core 2 Duo 2.66 GHz CPU, 8 GB memory, and Microsoft Windows 7 operating system was used. The software development tools used to implement the image-switching algorithm proposed in this paper are Visual C++ integrated development environment, MFC (Microsoft foundation class) library, and OpenCV [28]. The image database used for the experiment consists of various images captured by cameras in different types of indoor and outdoor circumstances without any specific constraint conditions.

The cameras used in this study are the Bumblebee 2 of Pointgrey [29] as a stereo camera, Kinect 2 of Microsoft as a TOF camera [15], and WMK-H302 of Wonwoo Engineering as an infrared camera [30]. In the different circumstances of illumination, edge, and distance, the proposed image-switching algorithm was applied to the experiment of this study. Figure 8 illustrates an example of the input image, the intensity image, the infrared image, and the depth map image used by the proposed algorithm.

Figure 8

An example of the images used.

Figures 9(a)–9(f) display an example of the images captured in different edge circumstances: “very plentiful,” “plentiful,” “normal,” “poor,” “very poor,” and “completely poor.” Figure 10 shows the edge images extracted from the input images in Figure 9.

Figure 9

Images with different edgeness.

Figure 10

Edge images.

Tables 2–4 present the overall camera types, which are dynamically chosen by the proposed image-switching algorithm based on the illumination, edge, and distance in the neighboring circumstances. In Tables 2–4, ST, TF, and IR stand for stereo camera, TOF camera, and infrared camera, respectively. It was difficult to present the illumination, edge, and distance on one three-dimensional table with three axes. Therefore, this paper presented three two-dimensional tables, each representing camera choice according to the combination of two features.

Table 2

Choice of camera based on illumination and edge.

Illumination	Edge
Illumination	Very plentiful	Plentiful	Normal	Poor	Very poor	Completely poor
Very bright	IR	IR	IR	IR	IR	IR
Bright	ST/IR	ST/IR	ST/IR	TF/IR	TF/IR	TF/IR
Normal	ST/IR	ST/IR	ST/IR	TF/IR	TF/IR	TF/IR
Dark	ST/IR	ST/IR	ST/IR	TF/IR	TF/IR	TF/IR
Very dark	TF/IR	TF/IR	TF/IR	TF/IR	TF/IR	TF/IR
Completely dark	IR	IR	IR	IR	IR	IR

Table 3

Choice of camera based on edge and distance.

Distance	Edge
Distance	Very plentiful	Plentiful	Normal	Poor	Very poor	Completely poor
Very near	IR	IR	IR	IR	IR	IR
Near	ST/IR	ST/IR	ST/IR	TF/IR	TF/IR	TF/IR
Normal	ST/IR	ST/IR	ST/IR	TF/IR	TF/IR	TF/IR
Distant	ST/IR	ST/IR	ST/IR	TF/IR	TF/IR	TF/IR
Very distant	IR	IR	IR	IR	IR	IR
Completely distant	IR	IR	IR	IR	IR	IR

Table 4

Choice of camera based on distance and illumination.

Distance	Illumination
Distance	Very bright	Bright	Normal	Dark	Very dark	Completely dark
Very near	IR	IR	IR	IR	IR	IR
Near	IR	ST/TF	ST/TF	ST/TF	TF	IR
Normal	IR	ST/TF	ST/TF	ST/TF	TF	IR
Distant	IR	ST/TF	ST/TF	ST/TF	TF	IR
Very distant	IR	IR	IR	IR	IR	IR
Completely distant	IR	IR	IR	IR	IR	IR

To judge whether the proposed image-switching algorithm is effective for actual situations, the proposed algorithm was applied to the detection of a moving object. That is, the proposed switching algorithm was used to detect a moving object through block-based motions and depth maps extracted from the captured images. Subsequently, the precision rate in (8) and the recall rate in (9) were applied to measure the accuracy of moving object detection on average according to each method:

\begin{matrix} R_{p r e c i s i o n} = \frac{N_{c}}{N_{c} + N_{f}}, \end{matrix}

(8)

\begin{matrix} R_{r e c a l l} = \frac{N_{c}}{N_{c} + N_{m}} . \end{matrix}

(9)

Figures 11-12 illustrate the graphs of precision and recall rates. As shown in Figures 11-12, the proposed method was better than the existing methods in terms of moving object detection. In particular, the proposed method chose the infrared camera to detect a moving object when the illumination was bright or dark and when an object was at a long distance. It used the stereo camera when the edge was plentiful and chose the TOF camera when the edge was poor. The process of detecting an object with the use of one fixed camera causes many errors when the neighboring circumstances are bad. Because the proposed method dynamically chooses the camera best suited to neighboring circumstances to detect a moving object, it offers better accuracy than other methods.

Figure 11

Precision rates.

Figure 12

Recall rates.

To sum up, the dynamic switching strategy proposed in this paper effectively integrates and applies the diverse features representing neighboring circumstances to reliably choose the one camera best suited to the current situation. In this sense, it is expected that the proposed algorithm will be useful when applied to a hybrid camera environment based on multiple cameras, which will be used often in the 2D or 3D computer vision and pattern recognition area.

6. Conclusion

To extract proper information from continuous input images more precisely in visual sensor networks, this paper proposed an image-switching strategy where, among different types of installed cameras, the one camera best suited to neighboring circumstances is chosen. The proposed strategy is to first receive initial images as input data and then extract multiple features representing neighboring circumstances from the input images. Subsequently, it is to define the neighboring circumstance metric, which is the sum of weights of the extracted features, and to dynamically switch cameras to obtain and analyze images in accordance with the neighboring circumstances.

According to the experiment to compare and evaluate the performance of the proposed method, the proposed dynamic switching strategy reliably chose, from among different types of cameras, the camera that was best suited to the neighboring circumstances. In particular, for dynamic switching, the proposed method applied a weighting factor to each feature to define an integrated metric, thus revealing that the method is not sensitive to changing neighboring circumstances and works robustly.

In future studies, the proposed dynamic switching algorithm should be applied to more types of cameras than the three types used in this paper. In addition, more research will be continuously conducted to obtain the robust features representing the neighboring circumstances.

Footnotes

Conflict of Interests

The authors declare that there is no conflict of interests regarding the publication of this paper.

Acknowledgment

This research was supported by Basic Science Research Program through the National Research Foundation of Korea (NRF) funded by the Ministry of Education, Science and Technology (2011-0021984).

References

Phamila

Y. A. V.

Amutha

Discrete cosine transform based fusion of multi-focus images for visual sensor networks

Signal Processing 2014 95 161 170

10.1016/j.sigpro.2013.09.001

Hooshmand

Soroushmehr

S. M. R.

Khadivi

Samavi

Shirani

Visual sensor network lifetime maximization by prioritized scheduling of nodes

Journal of Network and Computer Applications 2013 36 1 409 419

10.1016/j.jnca.2012.04.017

Chen

Y.-L.

Yen

H.-H.

MAC-aware and power-aware image aggregation scheme in wireless visual sensor networks

Journal of Sensors 2013 2013 13

414731

10.1155/2013/414731

2-s2.0-84875513956

Tian

Xiong

Design and implementation of the networked video surveillance and management platform in Suzhou subway line 1

Proceedings of the IEEE International Conference on Service Operations and Logistics, and Informatics (SOLI ′13)

July 2013

Dongguan, China

136 141

10.1109/soli.2013.6611397

2-s2.0-84886416947

Son

J.-Y.

Park

J.-H.

Moon

K.-D.

Lee

Y.-H.

Resource-aware smart home management system by constructing resource relation graph

IEEE Transactions on Consumer Electronics 2011 57 3 1112 1119

10.1109/tce.2011.6018863

2-s2.0-80052934601

Coşar

Çetin

Feature compression: a framework for multi-view multi-person tracking in visual sensor networks

Journal of Visual Communication and Image Representation 2014 25 5 864 873

10.1016/j.jvcir.2014.02.004

Semertzidis

Dimitropoulos

Koutsia

Grammalidis

Video sensor network for real-time traffic monitoring and surveillance

IET Intelligent Transport Systems 2010 4 2 103 112

10.1049/iet-its.2008.0092

2-s2.0-77953006388

Dornaika

Sappa

A. D.

A featureless and stochastic approach to on-board stereo vision system pose

Image and Vision Computing 2009 27 9 1382 1393

10.1016/j.imavis.2008.12.006

2-s2.0-67349084388

Schmalz

Forster

Schick

Angelopoulou

An endoscopic 3D scanner based on structured light

Medical Image Analysis 2012 16 5 1063 1072

10.1016/j.media.2012.04.001

2-s2.0-84861971993

10.

Cui

Schuon

Chan

Thrun

Theobalt

3D shape scanning with a time-of-flight camera

Proceedings of the IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR ′10)

June 2010

1173 1180

10.1109/cvpr.2010.5540082

2-s2.0-77955998137

11.

Fernández-Caballero

Castillo

J. C.

Martínez-Cantos

Martínez-Toms

Optical flow or image subtraction in human detection from infrared camera on mobile robot

Robotics and Autonomous Systems 2010 58 12 1273 1281

10.1016/j.robot.2010.06.002

2-s2.0-78649949405

12.

Light

Parthasarathy

McIver

Monitoring winter ice conditions using thermal imaging cameras equipped with infrared microbolometer sensors

Procedia Computer Science 2012 10 1158 1165

10.1016/j.procs.2012.06.166

13.

Liu

Wark

Lakemond

Sridharan

Self-calibration of wireless cameras with restricted degrees of freedom

Computer Vision and Image Understanding 2012 116 10 1033 1046

10.1016/j.cviu.2012.06.001

2-s2.0-84864047522

14.

Xiang

Yang

Liu

Zhou

Interfered depth map recovery with texture guidance for multiple structured light depth cameras

Signal Processing: Image Communication 2015 31 34 46

10.1016/j.image.2014.11.004

15.

C.-J.

Chiang

C.-Y.

Huang

J.-Y.

Kinect-enabled home-based rehabilitation system using Dynamic Time Warping and fuzzy logic

Applied Soft Computing Journal 2014 22 652 666

10.1016/j.asoc.2014.04.020

2-s2.0-84900700033

16.

Kakuta

Kondo

Ozaki

Arimoto

Yamada

Temperature imaging of sub-millimeter-thick water using a near-infrared camera

International Journal of Heat and Mass Transfer 2009 52 19-20 4221 4228

10.1016/j.ijheatmasstransfer.2009.04.024

2-s2.0-67650080151

17.

Amon

Hamins

Bryner

Rowe

Meaningful performance evaluation conditions for fire service thermal imaging cameras

Fire Safety Journal 2008 43 8 541 550

10.1016/j.firesaf.2007.12.006

2-s2.0-53149097359

18.

Cui

Jin

Jiang

A novel supervised feature extraction and classification fusion algorithm for land cover recognition of the off-land scenario

Neurocomputing 2014 140 77 83

10.1016/j.neucom.2014.03.034

2-s2.0-84901493073

19.

Jang

S.-W.

Pomplun

Kim

G.-Y.

Choi

H.-I.

Adaptive robust estimation of affine parameters from block motion vectors

Image and Vision Computing 2005 23 14 1250 1263

10.1016/j.imavis.2005.09.003

2-s2.0-28644431841

20.

Perez

Koch

Toward color image segmentation in analog VLSI: algorithm and hardware

International Journal of Computer Vision 1994 12 1 17 42

10.1007/bf01420983

2-s2.0-0028378715

21.

Yang

C. C.

Rodriguez

J. J.

Efficient luminance and saturation processing techniques for bypassing color coordinate transformations

Proceedings of the IEEE International Conference on Systems, Man and Cybernetics

October 1995

Vancouver, BC, Canada

667 672

2-s2.0-0029486907

22.

Beeran Kutty

Saaidin

Megat Yunus

P. N. A.

Abu Hassan

Evaluation of canny and sobel operator for logo edge detection

Proceedings of the International Symposium on Technology Management and Emerging Technologies (ISTMET '14)

May 2014

Bandung, Indonesia

153 156

10.1109/istmet.2014.6936497

23.

Azab

M. M.

Hussein

A. S.

Shedeed

H. A.

New technique for online object tracking-by-detection in video

IET Image Processing 2014 8 12 794 803

10.1049/iet-ipr.2014.0238

24.

Baha

Larabi

Accurate real-time neural disparity MAP estimation with FPGA

Pattern Recognition 2012 45 3 1195 1204

10.1016/j.patcog.2011.08.005

2-s2.0-80055025069

25.

Kim

S.-Y.

Lee

E.-K.

Y.-S.

Generation of ROI enhanced depth maps using stereoscopic cameras and a depth camera

IEEE Transactions on Broadcasting 2008 54 4 732 740

2-s2.0-57349140009

10.1109/TBC.2008.2002338

26.

Tsang

E. C. C.

Lee

J. W. T.

Yeung

D. S.

Tuning certainty factor and local weight of fuzzy production rules by using fuzzy neural network

IEEE Transactions on Systems, Man, and Cybernetics, Part B: Cybernetics 2002 32 1 91 98

10.1109/3477.979963

2-s2.0-0036465637

27.

Wang

Kong

Zhang

Maximum weight and minimum redundancy: a novel framework for feature subset selection

Pattern Recognition 2013 46 6 1616 1627

10.1016/j.patcog.2012.11.025

2-s2.0-84873170469

28.

Burden

Cleland

Conway

Falconer

Green

Chase

J. J. G.

Hann

Jermy

Palmer

Tracking a single cyclist during a team changeover on a velodrome track with Python and OpenCV

Procedia Engineering 2010 2 2 2931 2935

10.1016/j.proeng.2010.04.090

29.

Filipov

R. A.

Pádua

F. L. C.

Carone

M. A. B.

Pylon grid: a fast method for human head detection in range images

Neurocomputing 2013 100 74 85

10.1016/j.neucom.2011.12.040

2-s2.0-84868619099

30.

Woonwoo Engineering Company http://www.wonwoo.com