Working with OpenCV and Intel Image Processing Libraries.

Processing image Data Tools.

Faraón Llorens, Francisco José Mora , Mar Pujol, Ramón Rizo and Carlos Villagrá.

Grupo Vgia: Visión, Gráficos e Inteligencia Artificial

Departamento de Ciencia de la Computación e Inteligencia Artificial
Universidad de Alicante
E-03690, San Vicente, Spain
Phone +34 96 590 39 00, Fax +34 96590 39 02
e-mail: [email protected]

We will provide an overview of Intel OpenCV and Image http://groups.yahoo.com/ and then can subscribe by sending
Processing Libraries. We present an application of real-time email to [email protected].
gesture recognition using the libraries (segmenting a Intel has developed a uniquely decentralized research model
foreground object, creating Motion History Image (MHI), with over 80 labs situated around the world. The bulk of the
updating the intensity gradients, and recovering directional OpenCV software team resides Intel's Software Development
motion information). Center in Nihzniy Novgorod, Russia. Established in 1999, the
Some times most companies spent a lot of time and money center currently employees over 100 computer research
implementing those well-known techniques. OpenCV and scientists and engineers working in areas such as computer
IPL implement a huge amount of standard and advanced graphics, vision, media, Bayesian networks, compilers and tools
image processing techniques. With OpenCV we now have research. The support and new versions of the library are
access to a well implemented version that is fast and guaranteed.
reliable. OpenCV currently is supported to run on Windows and Linux
operating systems. We can use the library with programming
languages, for example Visual Basic or Java, and applications
like MathLab.
1. INTRODUCTION There are other libraries developed for the Intel group. For
example the Signal Processing Library is a library for work
OpenCV is computer-vision library for extracting and with audio signal, and Intel Vtune Performance analycer to
processing meaningful data from images. This library needs calculate the time-codes.
other Library, the Intel Image Processing Library. IPL provides IPL and OpenCV reference manuals detailed describe
a set of low-level image manipulation and OpenCV structure, operation ands functions of the libraries.
implements techniques and functions in the areas of
object/human/face segmentation, detection, recognition, and
tracking, as well as camera calibration, stereovision, and 2. INTEL IMAGE PROCESSING LIBRARY V2.5
2D/3D shape reconstruction.
OpenCV is an open-source release that includes C source The Intel Image Processing Library provides a set of low-level
code for all of the library’s functionality, it gives us confidence image manipulation functions in standard DLLs and static
in the code and we have know we can make local libraries form. The functions are optimized for Intel
modifications if necessary. Architecture processors, and are particularly effective at taking
Developed by the Intel research group, OpenCV is freely advantage of MMX™ technology, the Streaming SIMD
available at www.intel.com/research/mrl/research/opencv/ and Extensions (SSE) and SSE-2.. Currently, versions have been
has royalty free redistribution license. Anyone interested in developed for the Intel486™ and compatible processors, the
joining the user group needs to register with Yahoo groups at Pentium® processor, the Pentium Pro processor, the Pentium
processor with MMX technology, the Pentium processor, the
 3. THE OPENCV LIBRARY

OpenCv contains more than 500 funtions. While the API of

the library is C/C+, the bulk of the libray is made up of C
functions and the objects themselves are mostly self contained.
Figure 1. Some example areas would be Human-Computer
Interaction (HCI), Motion Tracking , Face Recognition,
Gesture Recognition and Segmentation and Recognition.

OpenCV is aimed at making computer vision accessible to

programmers and users in the area of real-time human-computer
interaction and mobile robotics. Thus, the library comes with
source code and hand-tuned assembly language binaries
optimized fot Intel processors, so that users can both leran from
the library and make use of its performance.
When you build and run an application using OpenCV, a built-
in DLL switcher is called at run time to automatically detect the
Pentium III processor, and, most recently, the Pentium 4
processor type and load the appropriate optimized DLL for that
processor. A separate DLL is available for each processor.
processor. If the processor type cannot be determined or if the
The library contains functions that perform filtering,
appropriated DLL is not available, an optimized C code DLL is
thresholding, and transforms (FFT, DCT, geometric), as well as
used. Included in the OpenCv download is the optimized Intel
arithmetic and morphological operations. The library uses a
Image Processing Library (IPL) on wich OpenCv partially
flexible image format, supporting channels of 1-, 8-, 16-, and
depends. Although it is included with OpenCV, you can also
32-bit integer pixels, as well as 32-bit floating point pixels, with
get IPL and other libraries for signal processing, matrix math,
an arbitrary number of channels per image. Conversion to and
JPEG and pattern recognition at
from the Windows* DIB (device independent bitmap) image
http://developer.intel.com/vtune/perflibst/.
format is supported, as are conversions between color and gray-
OpenCv ships with HTML overview and a detailed manual in
scale.
Pdf format. Added to this are many source-code examples,
Major changes from the Image Processing Library 1.0 include
detailed papers, and tutorials on differents topics. The source-
many more geometric functions, mask region of interest,
code examples currently include camera calibration, face
floating point support, example code, and color space
traking, kalman filter, condensation filter, face recognition,
conversions.
optical flow, and morphing and image to produce intermediate
Changes since version 2.2 include further optimization for
views. An interpretative C prototyping environment for
Pentium III processors and support for the Pentium 4 processor.
OpenCv and IPL is also available for download.
Additional information on this software as well as other Intel
Performance Libraries is available at
OpenCV addresses the areas of object/human/face
http://developer.intel.com/software/products/perflib/.
segmentation, detection, recognition, and tracking, as well as
camera calibration, stereovision, and 2D/3D shape
Intel® Image Processing Library (included in OpenCV
reconstruction. A full matrix algebra package is also included in
WinOS download):
the library to support algorithms in these areas.
• Image creation and access (same image header
Also included in the library are routines in:
used for both libraries).
• Image functions: Creation, allocation, destruction of
• Image arithmetic and logic operations.
images. Fast pixel access macros.
• Image filtering.
• Data Structures: Static types and dynamic storage.
• Linear image transformation.
• Contour Processing: Finding, displaying,
• Image morphology.
manipulation, and simplification of image contours.
• Color space conversion.
• Geometry: Line and ellipse fitting. Convex hull.
• Image histogram and thresholding.
Contour analysis.
• Geometric transformation (zoom-decimate,
• Features: 1st & 2nd Image Derivatives. Lines: Canny,
rotate, mirror, shear, warp, perspective
Hough. Corners: Finding, tracking.
transform, affine transform).
• Image Statistics: In region of interest: Count, Mean,
• Image moments.
STD, Min, Max, Norm, Moments, Hu Moments.
• Image Pyramids: Power of 2. Color/texture
segmentation.
 • Morphology: Erode, dilate, open, close. Gradient,
top-hat, black-hat.
• Background Differencing: Accumulate images and
squared images. Running averages.
• Distance Transform: Distance Transform
• Thresholding: Binary, inverse binary, truncated, to
zero, to zero inverse.
• Flood Fill: 4 and 8 connected
• Camera Calibration: Intrinsic and extrinsic,
Figure 2. Motion History Image for arm-streaching mvement
Rodrigues, un-distortion, Finding checkerboard
generated from image differences.
calibration pattern
• View Morphing: 8 point algorithm, Epipolar
alignment of images
Basically, directional motion information can be recovered
• Motion Templates: Overlaying silhouettes: motion
directly from the intensity gradients with the MHI. In addition,
history image, gradient and weighted global motion.
we provide a few simple motion features using these
• CAMSHIFT: Mean shift algorithm and variant
orientations. Figure 2 is a flowchart of this algorithm.
• Active Contours: Snakes
• Optical Flow: HS, L-K, BM and L-K in pyramid.
• Estimators: Kalman and Condensation.
• POSIT: 6DOF model based estimate from 1 2D view.
• Histogram (recognition) Manipulation,
comparison, backprojection. Earth Mover's Distance
(EMD).
• Gesture Recognition: Stereo based: Finding hand,
hand mask. Image homography, bounding box.
• Matrix: Matrix Math: SVD, inverse, cross-product,
Mahalanobis, eigen values and vectors. Perspective
projection.
• Eigen Objects: Calc Cov Matrix, Calc Eigen objects,
decomp. coeffs. Decomposition and projection.
• embedded HMMs: Create, destroy, observation
vectors, DCT, Viterbi Segmentation, training and test.
• Drawing Primatives: Line, rectangle, circle, ellipse, Figure 2: FlowChart algorithm.
polygon. Text on images.
• System Functions: Load optimized code. Get For this discussion, assume the camera is stationary. Althought
processor info. there are many more sophisticated techniques of segmenting a
• Utility: Abs difference. Template matching. Pixel foregroung object from the learned background, in this example
order<->Plane order. Convert Scale. Sampling lines. we label as potention foreground those pixel that are a set
Bi-linear interpolation. ArcTan, sqrt, inv-sqrt, number standard deviations from the mean color background.
reciprocal. CartToPolar, Exp, Log. Random numbs. When a new video image comes in, the foreground is separated
Set image. K-Means. from the background using background substraction. In order to
do this, we need to first learn the background model. The
learned background model will consist of the means and
4. WORKING WITH OPENCV AND IMAGE standard deviations of each pixel over many frames when no
PROCESSING LIBRARIES. foreground object is present. Listing 1 shows the routine that
does this. The pixel values and squared values of pixels are
We will present a real-time application that detects motions. assumed into floating-point images and these images are
This application has built on a standard Pc platform employing divided by the number of video frames (here 45 frames or 1.5
optimized OpenCv and IPL routines. seconds at 30 frames/sec) after collection is finished. This
For more details of this algorithm and code just described, go yields the mean and standard deviation of each pixel. Since
to “http://www.cse.lehigh.edu/FRAME/ IstdFP will be used as the threshold difference from Imean[x,y]
Davis/DavisBradski.htm” In listings you can se that all at wich a future pixel at x,y will be declared to be a foreground
functions and types of the OpenCV start with “cv” and Image pixel.
Processing Library with “ipl”. We construct a more localized
motion characterization for the Motion History Image that
extract motion orientations in real-time.
void captureBackground(IplImage //since we use Istd as a threshold,
*IvidIm, enforce that no threshold is too small
IplImage *IstdFP, IplImage float *pIstdFP = (float
*Iu, *)IstdFP->imageData;
IplImage *Istd) for(i=0: i<len: i++)
{ {
int height = IvidIm->height; if (*pIstdFP < 0.3) *pIstdFP = 0.3;
int width = IvidIm->width; pIstdFP++;
int I; }

//Create background mask (find mean and //meanI^2

variance of color background): iplMultiplySFP(IstdFP,
IplImage *Imean = IstdFP.backThresh);
iplCreateImageHeader(3. 0,
IPL_DEPTH_32F, "RGB", BGR", //convert to 8u images
IPL_DATA_ORDER_PIXEL, IPL ORIGIN_TL, convert32Fto8U (Imean, Iu);
IPL_ALIGN_QWORD,width, height, NULL, convert32Fto8U(IstdFP,Istd);
NULL, NULL, NULL); iplDeallocate(Imean, IPL_IMAGE ALL);
iplAllocateImageFP(Imean, 1,0.0);
int len = width*height*3; Listing 1.
Collect background mean and std of each pixel for background
//take statistics over 45 frames (~1.5 differencing code.
sees)
for(i = 0: i<45: i++)
{ After the mean and the standard deviation of the color
grabIm(IvidIm);//Get an image into background are calculated, the code then creates a binary
IvidIm mask of all possible regions are those pixels that deviate
cvAcc(IvidIm, Imean);//Accumulate it more than a given number of standard deviations from
into the mean values. Image dilation is used to help cose holes
// Imean in the foreground regions caused by image noise as in
cvSquareAcc(IvidIm, Listing 2.
IstdFP);//Accumulate
// squared image into IstdFP
}
//find mean and vars
//meanI
iplMultiplySFP(Imean, Imean,
(float)1.0/i);

//meanI^2
iplMultiplySFP(IstdFP, IstdFP.
(float)1.0/i); IplImage* ImeanSqr =
iplCloneImage(Imean);
iplSquare(ImeanSqr, ImeanSqr);

//Ivar = meanIA2 - (meanI)^2

iplSubtract(IstdFP, LneanSqr, IstdEP);
iplDeallocate(ImeanSqr, IPL_IMAGE_ALL);

//IstdFP = sqrt(Ivar)
cvbSqrt((const
float*)IstdFP->imageData,
(float*)IstdFP->imageData, len);
//Ii Video input image BGR cvCalcMotionGradient(IrhiIm, Imask,
//Im Mean image of background BGR Iorient, 3, MAX_TIME_DELTA,MIN
//Is Standard deviation of background from TIME_DELTA )
mean BGR
//Io Output image -- Grayscale //CALCULATE THE MOTION ORIENTATION
//Iot Temporary output image -- Grayscale globalDir = cvCalcGlobalOrientation (
//Itl,It2 Temporary images of same size, Iorient, Imask, ImhiIm,
depth and number of channels timestamp,MHI_DURATION):
BGR
//numIterations Number of dialations to globalDir = 360.0 - globalDir:
preform on foreground //Circularly rotate angle since TL
void backsubCVL(IplImage *Ii, IplImage *Im. origin
IplImage *Is, IplImage *Io, rather //than BL default
IplImage *Iot, IplImage *Itl, IplImage
*It2, int numIterations) Listing 3.
{ Update tMHi, Calculate the Motion Gradient and global motion
//Get 'Ii-Im orientation code.
iplSubtract(Ii,Im,Itl);
iplSubtract(Im,Ii,It2);
iplAdd(Itl,It2,Itl);

//Get Raw foreground = ~Ii-Im, > Is ?

255: 0;
iplSubtract(Itl,Is,Itl);
iplThreshold(Itl, Itl, 1);
iplColorToGray(Itl, Io);
iplThreshold ( Io , Iot , 1 );

//Fill up holes in the foreground

iplDilate(Iot,Io, numIterations);
} Figure 4. From tMHI to a gradient representation.
(a) tMHI (b) Gradients (c) Global Motion
Listing 2.
Extract the foreground and fill in holes code.
Table 1 list code timing for some of the IPL-OpenCv functions
invoved in this process. The timings are in form of clock cycles
per pixel, and speed-up factor is going from optimized C code
The foreground object is layered onto the “time Motion to optimized assembly is show in this column. Thus, We work
History Image” (tMHI). Image gradients of the tMHI are with 160x120 video images at 30Hz on 500MHz Pentium III.
calculated, the gradients directions indicate movement patters Intel´s Vtune Performance analycer code was used to record the
of the foreground object. In the OpenCv library, updatinf tMHI performance of the functions using both the optimised C (and
image is wrapeped into one function call, cvUpdate assembly versions of the libraries). Vtune rapidly samples the
MHIByTime(…). The image gradients are calculated from the application and produces a report aas to the precent of time
MHI via the call cvCalMotionGradient(…) this yields spent in each module.
directions of motion encoded implicity by the tMHI. Finally the
code computes the global direction of movement via the call
cvCalcglobalorientation(…). These calls are show in Listing 3. Function Optimized Optimized Speed-
Figure 4 depicts the process of going from the MHI, to a C MMX up
gradient representation to the global motion in that region. cvAcc 7.2 2.4 3.02
CvSquareAcc 9.9 2.6 3.77
//UPDATE THE MOTION HISTORY IMAGE CvSqrt 24.4 8.6 2.84
cvUpdateMHIByTime(IsilIm, IrhiIm, CvUpdateMHIByTime 13.0 8.5 1.63
timestamp, MHI DURATION): CvCalcMotionGradient 193.2 82.9 2.33
CvCalcglobalOrientation 23.7 21.3 1.11
//CREATE MOTION GRADIENT ORIENATIONS Table 1. Code Timing for some of the OpenCv functions
FROM THE tMHI
 7. ACKNOWLEDMENTS
5. USING OPENCV AND IPL WITH DIFERENT
OPERATING SYSTEMS AND PROGRAMMING This work has been supported by the Spanish Comisión
LANGUAGES Interministerial de Ciencia y Tecnología (CICYT),
project number TIC2001-0245-C02-02
IPL and OpenCV is supported to run on Windows
98/Millenium/NT/2000/XP. On Linux or other operating
systems the source code should build. Actually, There isn’t 8. REFERENCES
Official Linux Support.
[1] Intel Image Processinf Library Reference Manual. 2001
Objects and functions are Externed as C to avoid name
mangling, allowing OpenCV to be used with Visual Basic and [2] Intel Open Computer Vision Library Reference Manual.
Java. Object-oriented design is mostly intented to take place at 2001
COM or CORBa level for use in applications.
[3] Bradski, G. and Davis, James. “Real-time Motion
It’s easy and quickly work with IPL and OpenCV libraries in Template Gradientes using Intel CVLib”. ICCV'99
MsWindows with Microsoft Visual C++, Borland C++ and http://www.cse.lehigh.edu/FRAME/Davis/DavisBradski.ht
other C++ compilers. We can work with other programming m
languages but it’s more difficult, mainly in OPenCV. IPL can
be easily used with Borland Delphi to. [4] Bradski, G.,B-L Yeo, and M.Yeung. “Gesture for Video
content navigation” SPIE ’99, 356-24 s6, 1999.
In addition, the new 2.1 release includes an optional interface [5] Bradski, G. “Comeputer Vision Face Trackingfor use in
so all OpenCV functions can be imported into Matlab, one of perceptual user interface”. Intel Technology Journal.
the most widely used software development tools for computer http://developer.intel.com/technology/itj/q21998/articles/ar
vision research. t_2.htm. Q2. 1998
[6] P. Fieguth and D. Terzopoulos, “Color-based tracking
Hawk is an application distributed with OpenCV that allows of heads and other mobile objects at video frame rates,” In
write and run short scripts in C and see graphical results. It is Proc. Of IEEE CVPR, pp. 21-27, 1997.
application that integrates C iterpreter (modified version of EiC. [7] D. Comaniciu and P. Meer, “Robust Analysis of Feature
Original EiC site is located at http://www.kd-dev.com/~eic/), Spaces: Color Image Segmentation,” CVPR’97, pp. 750-
simple graphic library (highgui) and OpenCV+IPL (but, 755.
actually, any shared library can be attached to the application as [8] M. Kass, A. Witkin D.Terzopoulos, “Snakes: Active
plugin). contour Models,” Int. J. o f Computer Vision (1) #4, pp.
321-331, 1988.
We can work jointly with other libraries, for example
Microsoft Vision SDK or ImageMagick . Microsoft Vision
SDK is a library for writing programs to perform image
manipulation and analysis on computers running Microsoft
Windows operating systems. If users need to read and write
other file formats, it is possible to using a library named
ImageMagick

6. CONCLUSIONS

OpenCV and IPL implement a huge amount of standard and

advanced image processing techniques. It’s easy and cheap to
use these tools. IPL and OpenCV is supported to run on
Microsoft Windows and Linux operating Systems. We can use
these libraries with differents programming languages and tools
such as Microsoft Visual C++, Borland C++, Java, Borland
Delphi, Visual Basic, MathLab, etc… We quickly can use fast
and reliable implemented functions with IPL and OpenCV.