U.S. patent application number 09/798594 was filed with the patent office on 2001-07-19 for real-time object tracking system.
Invention is credited to Beach, Glenn J., Cohen, Charles J., Jacobus, Charles J., Paul, George V..
Application Number | 20010008561 09/798594 |
Document ID | / |
Family ID | 46257568 |
Filed Date | 2001-07-19 |
United States Patent
Application |
20010008561 |
Kind Code |
A1 |
Paul, George V. ; et
al. |
July 19, 2001 |
Real-time object tracking system
Abstract
A real-time computer vision system tracks one or more objects
moving in a scene using a target location technique which does not
involve searching. The imaging hardware includes a color camera,
frame grabber, and processor. The software consists of the
low-level image grabbing software and a tracking algorithm. The
system tracks objects based on the color, motion and/or shape of
the object in the image. A color matching function is used to
compute three measures of the target's probable location based on
the target color, shape and motion. The method then computes the
most probable location of the target using a weighting technique.
Once the system is running, a graphical user interface displays the
live image from the color camera on the computer screen. The
operator can then use the mouse to select a target for tracking.
The system will then keep track of the moving target in the scene
in real-time.
Inventors: |
Paul, George V.;
(Belleville, MI) ; Beach, Glenn J.; (Ypsilanti,
MI) ; Cohen, Charles J.; (Ann Arbor, MI) ;
Jacobus, Charles J.; (Ann Arbor, MI) |
Correspondence
Address: |
John G. Posa
Gifford, Krass, Groh
280 N. Old Woodward Ave., Suite 400
Birmingham
MI
48009
US
|
Family ID: |
46257568 |
Appl. No.: |
09/798594 |
Filed: |
March 2, 2001 |
Related U.S. Patent Documents
|
|
|
|
|
|
Application
Number |
Filing Date |
Patent Number |
|
|
09798594 |
Mar 2, 2001 |
|
|
|
09371460 |
Aug 10, 1999 |
|
|
|
60186474 |
Mar 2, 2000 |
|
|
|
Current U.S.
Class: |
382/103 ;
382/107; 382/165 |
Current CPC
Class: |
G06T 7/246 20170101;
G06V 40/20 20220101; A63F 2300/1093 20130101; G06V 40/28 20220101;
A63F 2300/6045 20130101; G06F 3/017 20130101; A63F 2300/69
20130101 |
Class at
Publication: |
382/103 ;
382/107; 382/165 |
International
Class: |
G06K 009/00 |
Claims
We claim:
1. A method of tracking a target, comprising the steps of:
inputting a sequence of images of a scene; selecting a target in
the scene; computing the center of the target in an initial one of
the images using one or more visual characteristics of the target
region; computing the center of the target in a subsequent one of
the images using the visual characteristics; and comparing the
center of the target in the subsequent image to the center of the
target in the initial image to determine movement of the target
within the scene.
2. The method of claim 1, wherein the visual characteristics
include the color, shape or the location of the target.
3. The method of claim 1, wherein the visual characteristics
include a combination of static and dynamic characteristics.
4. The method of claim 3, further including the step of modeling of
the dynamic characteristics to yield an estimate the location of
the target in the current image based on the location of the target
in previous images.
5. The method of claim 1, wherein the step of selecting a target in
the scene includes the step of user-selecting the target on a
computer screen through a graphical user interface.
6. The method of claim 5, wherein the graphical user interface
provides a bounding box surrounding the target superimposed on each
image as it is displayed on the screen.
7. The method of claim 2, wherein step of computing the center of
the target with respect to color further includes the steps of:
enabling a match between the color of the target in the subsequent
image to the color of the target in a previous image despite
differences arising from target lighting and shadows.
8. The method of claim 2, wherein step of computing the center of
the target with respect to color further includes the steps of:
enabling a match between the color of the target in the subsequent
image to the color of the target in a previous image within a
threshold of hue.
9. The method of claim 2, wherein step of comparing the center of
the target in the subsequent image to the center of the target in
the initial image includes a comparison of pixel in an RGB
format.
10. The method of claim 1, further including the step of
determining if the target has moved outside of the scene.
11. The method of claim 1, wherein: the visual characteristic is
color; and further including the step of finding a weighted average
of color to compute the center of the target based upon color
alone.
12. The method of claim 1, further including the step of segmented
a region defined by a predetermined closeness of color as an
estimate of target shape.
13. The method of claim 1, further including the step of continuing
to track the target when the target moves in front of or behind a
similarly colored object in the scene.
14. The method of claim 1, further including the step of continuing
to track the target when the target and input image move in
relation to one another.
Description
REFERENCE TO PRIOR APPLICATIONS
[0001] This application claims priority of U.S. provisional
application Serial No. 60/186,474, filed Mar. 2, 2000, and is a
continuation-in-part of U.S. patent application Serial No.
09/371,460, filed Aug. 10, 1999, the entire contents of each
application being incorporated herein by reference.
FIELD OF THE INVENTION
[0002] This invention relates generally to computer vision systems
and, in particular, to a real-time object tracking system and
method color involving a color matching technique requiring minimal
computation.
BACKGROUND OF THE INVENTION
[0003] Current imaging systems can convert live scenes into a
sequence of digital images which can be processed to track any
object in the scene from frame to frame. The techniques used for
tracking are numerous. Most of the currently available systems use
some characteristic of the subset of the image containing the
target to search and locate the target in the following image. The
quality and speed of the tracking system depends on the
implementation of this search and locate idea
[0004] Most tracking systems use correlation of a sample subimage
representing the object with parts of the current image. The
correlation values are computed in a search area around an
estimated location of the object. The correlation operation is
computationally expensive and usually is performed using
specialized hardware.
[0005] Another set of tracking methods uses a 3D model of the
object being tracked. In these methods, the model is mapped into
the target location based on the location and illumination
parameters. The disadvantage of such model based tracking methods
is the relatively high amount of computation for the mapping from
the 3D model to the image. The tracking systems that avoid the
correlation or model matching approaches, use characteristics of
the object's appearance or motion in estimating the location of the
object in the current image. These techniques are faster than
correlation methods but are less robust to changing shape and
temporary occlusion by similarly colored objects in the scene.
[0006] The work by Darell et al. in U.S. Pat. No. 6,188,777 uses
stereo cameras and involves three modules which compute the range
of the tracked object, segments the object based on color and does
pattern classification. Each of the modules involved places a large
computational load on the computer. The method of Peurach et. al.
in U.S. Pat. No. 6,173,066 uses a 3D object model database and
projection geometry to find the pose of the object in the 2D camera
image. The pose determination and tracking involves searching in a
multi-dimensional object pose space. The computation involved is
very high.
[0007] The method of Richards in U.S. Pat. No. 6,163,336 uses
special cameras and infrared lighting and a specialized background.
The method of Marques et. al. in U.S. Pat. No. 6,130,964 involves a
layered segmentation of the object in the scene based on a
homogenuity measure. The method also involves a high amount of
computation. The template matching method proposed by Holliman et.
al. in U.S. Pat. No. 6,075,557 which tracks subimages in the larger
camera image involves search and correlation means relatively large
amounts of computation. The method of Ponticos in U.S. Pat. No.
6,035,067 uses segmentation of the image based on pixel color. The
system of Wakitani in U.S. Pat. No. 6,031,568 uses hardware to do
template matching of the target. The method is computationally
expensive correlation is done via hardware.
[0008] The tracking proposed in this method by Suito et. al. in
U.S. Pat. No. 6,014,167 relies mostly on the difference image
between successive frames to detect motion and then tracks moving
pixels using color. This work uses correlation and searches in a
multi dimensional space to compute the object's 3D position and
orientation. The amount of computation involved is immense.
[0009] The proposed method of Matsumura et. al. in U.S. Pat. No.
6,002,428 does color matching to track the target The method of
Guthrie in U.S. Pat. No. 5,973,732 uses differencing and blob
analysis. The method of Hunke in U.S. Pat. No. 5,912,980 uses color
matching as opposed to shape. The method of Tang et. al in U.S.
Pat. No. 5,878,151 uses correlation to track subimages in the
image.
SUMMARY OF THE INVENTION
[0010] This invention resides in a real-time computer vision system
capable of tracking moving objects in a scene. Unlike current
search and locate algorithms, the subject algorithm uses a target
location technique which does not involve search. The system tracks
objects based on the color, motion and shape of the object in the
image. The tracking algorithm uses a unique color matching
technique which uses minimal computation. This color matching
function is used to compute three measures of the target's probable
location based on the target color, shape and motion. It then
computes the most probable location of the target using a weighting
technique. These techniques make the invention very computationally
efficient also makes it robust to noise, occlusion and rapid motion
of the target.
[0011] The imaging hardware of the real-time object tracking system
includes a color camera, a frame grabber, and a personal computer.
The software includes low-level image grabbing software and the
tracking algorithm. Once the application is running, a graphical
user interface displays the live image from the color camera on the
computer screen. The operator can then use the mouse to click on
the hand in the image to select a target for tracking. The system
will then keep track of the moving target in the scene in
real-time.
BRIEF DESCRIPTION OF THE DRAWINGS
[0012] FIG. 1 is a simplified drawing of an imaging system and
computer with tracking algorithm according to the invention;
[0013] FIG. 2 is a flow chart illustrating important steps of the
tracking algorithm;
[0014] FIG. 3 is a drawing of a preferred graphical user interface
for use with the system of the invention;
[0015] FIG. 4 is a series of drawings which show the use of color
to track a target or feature;
[0016] FIG. 5 illustrates the use a truncated cone to account for
slight variations in color; and
[0017] FIG. 6 illustrates steps of a method according to the
invention written in pseudocode.
DETAILED DESCRIPTION OF THE INVENTION
[0018] A schematic of the system is shown in FIG. 1. The imaging
hardware includes a color camera 102 and a digitizer. The sequence
of images of the scene is then fed to a computer 104 which runs
tracking software according to the invention. The tracking
algorithm is independent of the imaging system hardware. The
tracking system has a graphical user interface (GUI) to initialize
the target and show the tracking result on the screen 106.
[0019] The GUI for the ROTS displays a live color image from the
camera on the computer screen. The user can initialize the target
manually or automatically. Once initialized, the ROTS will then
track the target in real-time.
[0020] The flow chart of the tracking algorithm is shown in FIG. 2.
The program captures live images from the camera and displays them
on the screen. It then allows the user to select the target
manually using the mouse or automatically by moving the target to a
predetermined position in the scene. At the point of
initialization, the color, the shape and location of the target are
computed and stored. Once the target is initialized, we compute an
estimate of the target location using target dynamics. We then
compute the actual location using the color, shape and motion
information with respect to a region centered at the estimated
location.
[0021] The input to the ROTS is a sequence of color images,
preferably in the standard RGB24 format. Hence, the hardware can be
a camera with a image grabbing board or a USB camera connected to
the USB port of the computer. A preferred GUI is shown in FIG.
3.
[0022] Tracking using Color, Shape and Motion
[0023] Once the user clicks on the target in the image, we compute
the median color of a small region around this point in the image.
This will be the color of the target region being tracked in the
scene until it is reinitialized. We also store the shape of the
target by segmenting the object using its color. Once tracking
begins, we compute the center of the target region in the image
using a combination of three aspects of the target. The three
aspects are the color, the shape and the motion. This results in a
very robust tracking system which can withstand a variety of noise,
occlusion and rapid motion.
[0024] Color Matching
[0025] The color of a pixel in a color image is determined by the
values of the Red, Green and Blue bytes corresponding to the pixel
in the image buffer. This color value will form a point in the
three-dimensional RGB color space. When we compute the color of the
target, we assume that the target is fairly evenly colored and the
illumination stays relatively the same. The color of the target is
then the median RGB value of a sample set of pixels constituting
the target. When the target moves and the illumination changes the
color of the target is likely to change. We use a computationally
efficient color matching function which allows us to compute
whether a pixel color matches the target color within limits.
[0026] When the illumination on the target changes, the intensity
of the color will change. This will appear as a movement along the
RGB color vector as shown in FIG. 5. In order to account for slight
variations in the color, we further allow the point in color space
to lie within a small-truncated cone as shown in FIG. 5. The two
thresholds will decide the shape of the matching color cone. A
threshold on the angle of the color cone and another threshold on
the minimum length of the color vector decides the matching color
space. Thus, any pixel whose color lies within the truncated cone
in color space will be considered as having the same color as the
target.
[0027] Given a colored pixel, we quantitatively define the match
between it and a reference color pixel as follows. Let (R, G, B) be
the values of the RGB vector of the first pixel. Let (R.sub.r,
G.sub.r, B.sub.r) be the RGB vector for the reference color. 1 d =
RR r + GG r + BB r m r = R r 2 + G r 2 + B r 2 m = R 2 + G 2 + B 2
d m = d m r d 1 = d m r m ColorMatch ( R , G , B ) = { d m d a if (
( d m l < d m < d m h ) & ( d 1 l < d a < d a h ) )
0 otherwise
[0028] The value of d.sub.m is related to the length of the
projection of the given color vector onto the reference vector. The
value of d.sub.a is related to the angle between the two vectors.
If we set two threshold bands for d.sub.m and d.sub.a, we can
filter out those pixels which lie within the truncated cone around
the reference vector. Their product will indicate the goodness of
the match. The parameters d.sub.m and d.sub.a are chosen to be
computationally simple to implement which becomes important when
all the pixels in a region have to be compared to the reference
color in each new image.
[0029] Position Using Color
[0030] Once we have the target color and a color matching
algorithm, we can find all the pixels in any given region of the
image which match the target color. We use the quantitative measure
of the match to find a weighted average of these pixel positions.
This gives us the most likely center of the target based on color
alone. If (i, j) are the row and column coordinates of the pixel
P.sub.c(i,j), then for a given rectangular region the most likely
target center based on color alone will be given as follows. 2 P c
( i , j , t ) = ColorMatch ( R ( i , j , t ) , G ( i , j , t ) , B
( i , j , t ) ) Center color = [ r c c c ] = [ 1 I * J P c ( i , j
, t ) * i 1 I * J P c ( i , j , t ) 1 I * J P c ( i , j , t ) * j 1
I * J P c ( i , j , t ) ]
[0031] Note that the centroid of the target is computed as a
weighted sum. The weights are the color matching measure of the
pixel. This weighting of the pixel contrasts with the usual
practice of weighting all matching pixels the same makes our
algorithm less prone to creep. We also keep track of the sum of the
matched pixel weights. If this sum is less than a threshold we
assume that the target is not in the region.
[0032] Shape Matching
[0033] Once the target is initialized, we compute a two-dimensional
template of the target. We use this dynamic template which is
updated every frame to measure the closeness of pixels at the
estimated location to the target shape. Given the color of the
object being tracked and the color matching function we segment all
the pixels in a region around the estimated location. The resulting
segmented image is the shape of the object and forms the template.
With each new image of the scene, the template of the target in the
previous frame is used to compute the new center of the target in
the new image. The advantage of using templates instead of any
assumed shape such as an ellipse is that the tracking and
localization of the target is much more robust to shape change and
hence more accurate. 3 P ( i , j , t ) = ColorMatch ( R ( i , j , t
) , G ( i , j , t ) , B ( i , j , t ) ) for time = t M ( i , j , t
- 1 ) = { 1 if ( P ( i , j , t - 1 ) > 0 ) 0 otherwise S ( i , j
, t ) = P ( i , j , t ) M ( i , j , t - 1 ) Center shape = [ r s c
s ] = [ 1 I * J S ( i , j , t ) * i 1 I * J S ( i , j , t ) 1 I * J
S ( i , j , t ) * j 1 I * J S ( i , j , t ) ]
[0034] The closeness of the shape is a summation of the product of
the pixel color match P(i, j) with the target template M(i, j).
Note again that the color matching measure is used to weight the
shape measure. This makes our algorithm robust to creep. Once the
region S is obtained, we can compute the centroid of S. This is the
probable location of the target based solely on the shape of the
target.
[0035] Motion Detection
[0036] The algorithm checks for motion in a region near the
estimated target position using a motion detecting function. This
function computes the difference between the current image and the
previous image, which is stored in memory. If motion has occurred,
there will be sufficient change in the intensities in the region.
The motion detection function will trigger if a sufficient number
of pixels change intensity by a certain threshold value. This
detection phase eliminates unnecessary computation when the object
is stationary.
[0037] Position Using Motion
[0038] If the motion detection function detects motion, the next
step is to locate the target. This is done using the difference
image and the target color. When an object moves between frames in
a relatively stationary background, the color of the pixels changes
between frames near the target (unless the target and the
background are of the same color). We compute the color change
between frames for pixels near the target location. The pixels
whose color changes beyond a threshold make up the difference
image. Note that the difference image will have areas, which are
complementary. The pixels where the object used to be will
complement those pixels where the object is at now. If we separate
these pixels using the color of the target, we can compute the new
location of the target. The set of pixels in the difference image,
which has the color of the target in the new image, will correspond
to the leading edge of the target in the new image. If we assume
that the shape of the target changes negligibly between frames, we
can use the shape of the target from the previous image to compute
the position of the center of the target from this difference
image.
[0039] Let D be the difference sub-image between the previous
target and the estimated target location in the new image. If we
threshold the difference image, we end up with a binary image. If
we intersect this binary image D with the shape of the target in
the new image M we get the moving edge of the target as the region
V. We then weight this region by the color matching measure P. 4 D
( i , j , t ) = { 1 if ( P ( i , j , t - 1 ) - P ( i , j , t - 1 )
> m 0 otherwise M ( i , j , t ) = { 1 if ( P ( i , j , t ) >
c ) 0 otherwise V ( i , j , t ) = ( D ( i , j , t ) M ( i , j , t )
) * P ( i , j , t ) Center motion = [ r m c m ] = [ 1 I * J V ( i ,
j , t ) * i 1 I * J V ( i , j , t ) 1 I * J V ( i , j , t ) * j 1 I
* J V ( i , j , t ) ]
[0040] The centroid of the region V is then computed as the
probable location of the target based on motion alone. This
weighting of the intesection region by the color matching measure
makes out tracking less prone to jitter.
[0041] In a physically implemented system, the image capture board
is capable of providing us with a 480.times.640-pixel color image
at 30 frames per second. Processing such a large image will slow
down the program. Fortunately, the nature of the tracking task is
such that, only a fraction of the image is of interest. This region
called the window of interest lies around the estimated position of
the target in the new image. We can compute the location of the
target in t he new image from the location of the target in the
previous image and its dynamics. We have used prediction based on
velocity computation between frames. This technique is able to keep
track of the target even when the target moves rapidly. We have
found that the window of interest is typically one one-hundredth
the area of the original image. This speeds up the computation of
the new target location considerably.
[0042] Tracking Algorithm
[0043] If we are given an estimated target location as (rc, cc) in
the new image and the size of the area to be searched is given by
(rs, cs), then the algorithm can be written in pseudo code as shown
in FIG. 6.
[0044] Note that the color matching weight c is being used to
weight all the three centers. This weighting makes this algorithm
smoother and more robust. The velocity computed at the end of the
tracking algorithm is used to compute the estimated position of the
target in the next frame.
[0045] Extensions of the system are possible in accordance with the
described algorithm herein. One is a tracking system which can
track multiple targets in the same image. Another uses the tracking
in two stereo images to track the target in 3D.
* * * * *