Finger Tracking for Breast Palpation Quantification
With Stereo Color Cameras
Jianchao Zeng[1], Yue Wang[2], Matthew Freedman and Seong K. Mun
Imaging Science and Information Systems Center
Department of Radiology, Georgetown University Medical Center
2115 Wisconsin Avenue, NW, Suite 603, Washington, D.C., 20007
phones:
202-687-1533 (zeng), 7953 (wang), 7948 (freedman), 7955 (mun)
fax:
202-784-3479
emails:
[zeng|yuewang|freedman|skm]@isis.imac.georgetown.edu
ABSTRACT
Early detection of breast cancer, one of the leading causes of death by cancer for women in the United States is key to any strategy designed to reduce breast cancer mortality. Breast self-examination (BSE) is considered as the most cost-effective approach available for early breast cancer detection because it is simple and non-invasive, and a large fraction of breast cancers are actually found by patients using this technique today. In BSE, the patient should use a proper search strategy to cover the whole breast region in order to detect all possible tumors. At present there is no objective approach or clinical data to evaluate the effectiveness of a particular BSE strategy. Even if a particular strategy is determined to be the most effective, training women to use it is still difficult because there is no objective way for them to know whether they are doing it correctly. We have developed a system using vision-based motion tracking technology to gather quantitative data about the breast palpation process for analysis of the BSE technique. By tracking position of the fingers, the system can provide the first objective quantitative data about the BSE process, and thus can improve our knowledge of the technique and help analyze its effectiveness. By visually displaying all the touched position information to the patient as the BSE is being conducted, the system can provide interactive feedback to the patient and create a prototype for a computer-based BSE training system. We propose to use color features (e.g., color tapes or nail polish), put them on the finger nails and track these features, because in breast palpation the background is the breast itself which is similar to the hand in color. This situation can hinder the ability/efficiency of other features if real time performance is required. To simplify feature extraction process, color transform is utilized instead of RGB values. Although the clinical environment will be well illuminated, normalization of color attributes is applied to compensate for minor changes in illumination. Neighbor search is employed to ensure real time performance, and a three-finger pattern topology is always checked for extracted features to avoid any possible false features. After detecting the features in the images, 3D position parameters of the colored fingers are calculated using the stereo vision principle. In the experiments, a 15 frames/second performance is obtained using an image size of 160x120 and an SGI Indy MIPS R4000 workstation. The system is robust and accurate, which confirms the performance and effectiveness of the proposed approach. The system can be used to quantify search strategy of the palpation and its documentation. With real-time visual feedback, it can be used to train both patients and new physicians to improve their performance of palpation and thus improve the rate of breast tumor detection.
Subject terms: Vision-based finger tracking, color feature extraction, color transform, 3D position calculation by stereo vision, medical applications, breast palpation for cancer detection, real-time visual feedback
1. Introduction
1.1 Methods of
detecting breast cancer
Mammography
is acknowledged to be the best method for the detection of early breast cancer,
but despite the availability of this excellent method many breast cancers are
first found by palpation. Breast examination is known to be an important
supplement to mammography capable of allowing detection of cancers that
mammography fails to identify. The patients in whom breast palpation is most
important are: (1) the younger women with radiodense breasts (in whom
mammography is less accurate) and (2) women who have rapidly growing tumors
that develop between mammographic examinations. Mammography is best for finding
small tumors manifest by microcalcifications and all tumors in fatty replaced
breasts. Small masses in the glandular breasts of (mainly) younger women are
harder to find and breast palpation is of use in these women. Approximately 50%
of cancers do not contain calcifications. Their appearance when small resembles
that of normal breast tissue and of small benign breast nodules. On palpation,
however, these small tissue densities are harder than normal tissue, allowing
their detection.
1.2 Interval
cancers as an indicator of the value of breast physical examination
One can
gain an estimate of the value of breast physical examination by looking at the
interval cancer rate in mammography screening trials. In the Malmo study with
an 18 to 24 month interval between mammograms1, 17% of the cancers
found were cancers found during the interval between mammography screening. In
women age 45-49, 27 % were interval cancers. 29% were interval cancers in ages
50-54. Between ages 65-69, interval cancers were 17%. In the Swedish Two County
Study with a 24 month interval between mammograms2, 50% of the
cancers (not including DCIS) detected between ages 40-49 were interval cancers.
This compares to 38% of interval cancers ages 50-59. The method of detection of
the interval cancers in these two studies is not specified, but it would have
been based on symptoms leading to further evaluation, clinical breast
examination (CBE) or breast self-examination. The interval cancer rates suggest
strongly that breast physical examination performed either as BSE or CBE with
or without symptoms is a major method for the detection of breast cancer in
younger women when the mammogram interval is 18-24 months.
Sickles
estimates their interval cancer detection rate as 7%3; however,
there has been no statewide cancer registry for him to use to be certain that
this estimate is accurate. Because only 27% of the women in his series have had
a follow-up mammogram by his group, this false negative or interval cancer rate
may be underestimated.
Tumors
appearing in the intervals between mammography screenings are of higher histologic
grade and larger size than those detected by mammography2. Ages
40-49 had 35% grade 1 in screening detected cancers vs. 13% in interval cancers
and the screening detected cancers were an average of 11.7 mm vs. 16.7 mm in
the interval cancers. Grade 2 were 27% vs. 35% and grade 3 were 38% vs. 52 %.
Tumors
found by CBE are, on average, larger than those found by mammography. Tumors
found by BSE tend to be larger than those found by either CBE or mammography
and the long term survival is less. Studies of BSE do, however, show that many
smaller and still curable cancers are found.
1.3 Survival from
breast cancer depending on method of initial identification
If one
looks at long term survival statistics of women with breast cancer based on the
method of tumor identification, those women whose tumors are found by
mammography or clinical breast examination do best with a ten year survival of
77% and 78%. Those found by breast self-examination do less well with a ten
year survival of 64%. Conversely, in this same series (which was done prior to
the availability of widescale screening mammography), 82% of women
self-detected their cancers. Tumor size was also directly related to method of
detection. The average size of tumors detected was 1.4 cm by mammography, 2.1
cm by clinical breast examination and 2.7 cm by self detection. Axillary node
metastases were more often present when the cancer was found by self detection
(44%) compared to CBE (27%) and mammography (10%).
1.4 Effectiveness
of training in breast self-examination
Cancers
identified by practicers of BSE have a better long term prognosis than those
found by sporadic methods.
Study showing improved survival: Le Geyte4,
in his study of 616 women with breast cancer reported on 226 who had been
taught BSE and 390 who had not. The six year survival rates were 73.1% in the
BSE taught group and 66.1% in those untaught.
Studies using intermediate outcomes:
Reported studies, with shorter clinical
follow-up, used intermediate outcome measures of tumor size and stage. These
show that women who are trained in and use BSE do find smaller tumors of lower
stage and should therefore have better long term survival than those who do not
use BSE.
Mant
reported that those practicing BSE identified smaller tumors at lower stage5.
Tumors less than or equal to 2 cm were 45% in those who practiced BSE vs. 33%
(without BSE) . Clinical stage T-1 was 42% (with BSE) vs. 27% (without BSE).
N-0 pathological stage was 50% (with BSE) vs. 37% (without BSE).
Semiglazov
reported the initial results of a randomized prospective World Health
Organization sponsored study of 31,186 women trained in BSE compared to 31,066
who were not trained6. He reported that those trained identified
smaller tumors and that these were at a lower stage. The BSE trained group had
28% of their tumor identified at less than 2 cm (T-1) in size compared to only
19% in those not trained. Both groups had an N-0 of 44%, but the BSE trained
group had a 33% N-1, compared to 25% for those not trained and a 17% N-2
compared to 31% N-2 for those not trained.
Training
in BSE has been clearly documented to allow women to find smaller breast
cancers and to have a better long term survival than cancers found by sporadic
examination methods. The survival of women with cancers found by BSE is much
greater than that of women whose methods are sporadic.
1.5 Effectiveness
of different methods of training in BSE
The
effectiveness of different methods of training has been evaluated by Fletcher7.
She compared three main methods of training, each with and without physician
encouragement. She found that those trained with nurse instruction and
semi-realistic models performed better than those trained by nurse instruction
without models and those who were given the American Cancer Society pamphlet on
BSE and who received encouragement to practice BSE. One year later, those trained with the models were better at
detecting lumps (57%) in the models than those trained without the models (47%)
and those who had received only the pamphlet and encouragement to do BSE (45%).
It has
been shown, therefore that BSE is an effective method of finding curable breast
cancer, that those who have been trained in this method find smaller cancers at
an earlier stage in their own breasts than those who have not been trained and
that training with models improves a woman’s skill in finding lumps in the
models when measured one year following the training.
1.6 Making better
models
The
methods of training in BSE are important in the value of the procedure. In the
study reported by Semiglazov6, instructors met with groups of 5 to
20 women of whom one was used to demonstrate the technique. BSE was found to
improve the early detection of breast cancer compared to those women not trained.
Fletcher showed that providers and women trained using the models made by
Mammatech Corporation were better in detecting lumps in these models after
training compared to those who received training by a nurse who did not use the
models7.
The
following problems are present in the current training methods. (1) Current
models are based on clinical cases. They may necessarily represent a
standardization process, but they are not individualized to the patient’s own
breast composition. (2) Training women in BSE requires continuous supervision
by a trainer. We believe that improved methods will allow some of this training
to be performed by a computer and that this will improve both the quality and
availability of training. (3) Important issues such as search strategy and full
breast coverage are evaluated subjectively. Quantitative assessment will
greatly improve both the rate and accuracy of detection. (4) While the pattern
of her hand motion can be observed by the trainer, there is no indication to
the woman or the trainer that a proper amount of pressure has been applied. (5)
Lumps are felt only in the plastic model, not in the woman’s own breast.
We are
developing an integrated system to solve these problems. We consider it
important that a woman have visual feedback of where her fingers have been
during her training in BSE. By projecting the finger tracking camera on her,
she can observe herself and see recorded regions she has reached or not
reached. One can also set the finger tracking system into test mode, let the
woman palpate her breasts without viewing the computer record and then review
where the fingers have been and the pattern of palpation followed by
reactivating the computer.
The force
applied during breast palpation is also important. The models we are making
will allow continuous feedback of the force and direction of finger motion. We
will set these to indicate the proper zone of pressure so that the woman will
be able to see on the computer screen whether her palpation is too soft, too hard
or within the proper range.
Each
woman’s breast has a somewhat different texture, thus no plastic model is
optimum. The long term goal of our
system is that it could produce, within a woman’s own breast, the
sensation of a mass that is of appropriate hardness and size. We can thereby
personalize the training model to each woman.
1.7 Stereo vision-based finger tracking
Clearly, traditional approaches to BSE are featured by their subjective nature. For BSE to be effective, the patient must use a proper search strategy to cover the entire breast region, but there is currently no objective clinical data in evaluating the effectiveness of a particular BSE strategy. Even if a particular strategy is determined to be the most effective, training a woman to use it is still difficult because there is no precise way for her to determine if she is doing correctly. In this paper, we propose a stereo vision-based motion tracking technology to gather quantitative finger-position data and provide real-time visual feedback during the breast palpation process. By tracking the position of the fingers, the system could provide the first-hand objective quantitative data about the BSE process, which can improve our knowledge of the technique and help analyze its effectiveness. By displaying position information to the patient as the BSE is performed, the system could be used to give interactive feedback to the patient and create a prototype for a computer-based BSE training system.
While there exist other tracking technologies (e.g. magnetic, acoustic) which could be used to track the hand motion, vision-based tracking is considered the most appropriate for a practical BSE data acquisition system because it is the least obtrusive and least expensive technique. Vision-based hand tracking is under investigation by many researchers8-11. Most of them use model-based method, in which 3D or even 2D models of a generic human hand are employed and fitted to the specific hand of a user for the tracking and recognition of 3D hand gestures. These methods are generally slow for real-time data acquisition so, for the specific situation of breast palpation, we propose a color-assisted finger tracking metaphor in this paper, which tracks the 3D spatial positions of colored finger nails. Color transform is utilized in color feature extraction, instead of directly using RGB values. Normalization of color attributes is used to tackle the problem of any possible minor ambient lighting variations. The relatively unchanging three-finger nail pattern is employed to differentiate the target fingers from any possible false patterns in the background. A pair of cameras are employed for stereo depth calculation, and the real-time performance can be achieved without using any special hardware.
2. Color-based tracking approach
2.1 Comparison of color transforms
In the situation of breast palpation the background is the breast itself which is very similar to the hand in color. This special situation makes ordinary feature extraction techniques, such as edge detection, very difficult to work effectively since real-time performance is also required. Artificial makers are therefore considered more appropriate for this situation. Specially shaped geometric markers, however, are difficult to be applied since the fingers are too small in size to put these markers on, and they are not planar and therefore will cause undesirable geometric deformation to the markers. As a result, we propose to put special color features (such as color tape, color finger polish) on top of the finger nails. This color approach is also expected to be advantageous in real-time performance, compared to geometric gray-scale marker approach which needs to detect edges and infer shape information.
There are many color coordinate systems, such as RGB, HSI, LHS, YIQ etc. Each of them has its own advantages and disadvantages, and therefore they are selected and used according to the special requirements of the actual applications12-14. Basically, the color of a pixel in an image is initially represented as a vector of red (R), green (G) and blue (B) values. These values can be transformed in different color spaces like HSI, YIQ to get such values of the pixel as hue, saturation and luminance. Some of these transforms are nonlinear (e.g., HSI, LHS) and the others are linear (e.g., YIQ, XYZ). In practical applications, linear transforms are often superior to nonlinear ones because nonlinear transforms can result in some unexpected singularities15. In addition, if real-time performance is required, linear transforms are even more preferred because of their simplicities.
Unfortunately, color images captured by a camera are largely affected by the environmental lighting and shadowing conditions. The original color of an object is easily “hidden” by such factors as strong highlights and shadows, and therefore even the above-mentioned color transforms often have difficulties in removing these factors to get the real color values. This has invoked research into impacts of physical processes during the formation of an image on captured color properties, and recently color extraction and segmentation based on physical reflection models have been proposed, which can remove highlights and other factors from the image and which have shown better results16, 17.
However, these methods are still in their early stage and are generally computation intensive, and therefore they are not often employed in actual applications compared to those approaches based on color transforms. Here we also make use of color transform and, after an experimental comparison of various coordinate systems, we select YIQ color space for color feature extraction in our system. We have compared three color spaces: (H, S, I), (I1, I2, I3), (Y, I, Q) which are defined as below.
(1)
(2)
(3)
where 
.
Fig. 1 shows an example of the effects of the color transforms. It can be seen that (H, S, I) is not robust to the noises in this situation, and both (Y, I, Q) and (I1, I2, I3) can obtain satisfactory results. However, based on our experiments, (I1, I2, I3) transform is more sensitive to highlights than (Y, I, Q).
(a) Input image (b) (I1, I2, I3) transformed image
(c) (Y, I, Q) transformed image (d) (H, S, I) transformed image
Fig. 1 Comparison of different color transforms
2.2 Color feature extraction and grouping
2.2.1 Feature extraction
We have considered two methods to extract color features: a template matching method and a threshold-based extraction, and compared their performance in the experiments. In template matching, we have selected the three colored finger nails as a unit template pattern, assuming that this pattern will not change much for the same user during the process of palpation. By successfully matching the real input image, this method can conduct color feature extraction and grouping at the same time. To deal with lighting changes, we have calculated the Hamming distances by using only the I and Q components in each pixel.
In threshold-based color feature extraction, we have made use of the following two values as discriminants, corresponding to the hue and saturation values of a pixel, respectively, and we have used multiple empirical thresholds for these two values with respect to different Y values.
h
= arctan(Q/I), s = sqrt(I*I+Q*Q)
We have implemented both methods in the experiments for color feature extraction, and we found that the threshold-based method is more robust to the environmental noises such as shadowing. By employing the neighbor search technique, the threshold-based method can also maintain a speed advantage over the template matching method. Therefore, the threshold-based method is selected in our experimental system.
Although the environment is supposed to be well-illuminated, effect of minor changes in illumination and other noises is still considered in the system. First, normalization of (R, G, B) vectors is performed before color transform. Then, a noise removal algorithm is implemented as follows.
·
Do color
transform for the breast-only image (background image)
·
Extract color
features using the threshold-based method, and binarize the image and denote it
as I1(i, j)
·
Capture a frame
of palpation and calculate its binary image in the same way, and denote it as
I2(i, j)
Create a binary image F(i, j) as follows:
This noise removal algorithm is quite effective in removing noises caused by lighting changes, such as shadows. This image F(i, j) will be used as input to the feature grouping algorithm.
2.2.2 Feature grouping
The grouping algorithm consists of three steps of “group formation”, “group verification” and “three-finger pattern checking”. It is primarily based on criteria of distance between pixels, pixel numbers, pixel centralization degree and pixel group radius. In group formation, if the distance between two pixels is larger than a default value, they are classified into different groups. In group verification, if a group has pixels of less than a default number or more than some constraint number, or if its centralization degree CD (defined below) is less than a threshold, or if its radius GR (also defined below) is beyond a pre-defined scope, the group is regarded either as isolated noise or as non-finger region and therefore discarded. After verification, the three-finger pattern topology is checked for each triple groups. This topology is in a small near-isosceles-triangular pattern among the centroids of the three finger feature groups, and each pair of the centroids should satisfy a distance constraint. The grouping is outlined below:
(1) Group formation
for
all the extracted feature pixels Pn
if (||Pi-Pj|| < D1)
then Pi, Pj -> Gi
else Pi -> Gi & Pj -> Gj
endif
endfor
(2) Group verification
for
all the formed groups Gm
if (N1 < N(Gi) < N2 && CDi
> Deg && R1 < GRi < R2)
then Gi -> set(G)
endif
endfor
(3) Three-finger pattern checking
for
all the groups in set(G)
if (D2 < ||Gi-Gj|| < D3)
then ||Gi-Gj|| -> set(Distance)
endif
endfor
if (most-similar(||Gi-Gj||,
||Gi-Gk||)) /* find two most similar
distances in set(Distance) */
then
if (not-on-line(Gi, Gj, Gk)) /* check if the three groups are not on a
line */
then (Gi, Gj, Gk) -> goal-posi /* accepted as final finger positions */
endif
endif
where, ||.|| stands for calculation of distance, N(.) for calculation of number, set(.) for a set of objects such as groups and distances. D1, D2, D3, N1, N2, Deg, R1 and R2 are default values. CDi = N(Gi)/(number of connected areas in Gi), GRi = max(//Pi-Pj//) for pixels in Gi. The grouping process is shown in Fig. 2.
(a) Extracted features (b) Group formation
(c) Group verification and pattern checking (d) Final finger positions
Fig. 2 Feature extraction and grouping
2.3 Position calculation by stereo vision
After detecting and grouping the features in the images, 3D position parameters of the colored fingers are calculated. In our experimental environment, the origin of a world coordinate system is set at the lens center of the left camera, and the x-axis is set across the two lens centers. The z-axis is set to coincide with the optical axis of the left camera. Suppose P(x, y, z) is the center point of a feature nail, and its projective projections on both the left and right camera images are (xl, yl) and (xr, yr), respectively, which are measured from image planes. Let d be the distance of the two cameras, and f be the focal length of the cameras, then the 3D position of P is calculated as follows.
Because the finger features in both left and right images are clear and uniquely defined, there is no difficulty in finding correspondence among these features. f and d can be determined through calibration process.
Fig. 3 Setup of the developed palpation system
(a) An example of real-time finger nail tracking (b) Standard search pattern for breast palpation
(c) Visual feedback of the search pattern (d) Visual feedback of the coverage
Fig. 4 An experimental example of finger tracking and visual feedback
3. Experimental results
We have conducted experiments of finger tracking using the proposed approach, and a 15 frames/s performance is obtained using an image size of 160x120 and an Indy MIPS R4000 workstation. The system setup is shown in Figure 3. Figure 4 shows an example of the experiments. Fig. 4 (a) shows one frame of the real-time finger tracking during palpation, Fig. 4 (b) gives a standard search pattern as a reference guidance for the user. Fig. 4 (c) is the real-time visual feedback of search pattern during a real palpation process, where only the middle finger positions are displayed, while Fig 4 (d) gives the real-time visual feedback of entire three-finger coverage on the breast during a similar palpation process. The system is quite robust and accurate in the experiments, which confirms the performance and effectiveness of the proposed tracking approach.
4. Discussions
In implementing the tracking technique and the system, the following issues have been discussed.
(1) Tracking accuracy
The system can track a finger feature in an accuracy of about 3 mm (a pixel) when the cameras are placed at about 400 mm from the breast , which is much smaller than a finger itself.
(2) Highlights and shadows
To some extent, the system can deal with the noises in the input image caused by highlights and shadows. However, if the noises are too strong, the system may fail to extract correct features. Although this problem can be avoided by controlling the environmental lighting condition, the final solution may depend on progress in physical reflection models research.
(3) Feature grouping
In some rare situations, the colored finger nails may appear connected to one another in the image and therefore may cause some difficulties in feature grouping. With current grouping algorithm, the system may make mistakes in calculating the finger positions in such situations. This problem can be partially solved by incorporating axial projection procedures in the grouping algorithm, which is now under development.
(4) Obstruction
Because we are using a pair of cameras in tracking, total obstruction of both cameras is almost impossible. However, when either camera is obstructed, depth information cannot be calculated properly. In such case, depth information will be estimated using projective scale changes.
(5) Interactive feedback
We are working toward two kinds of feedback. The first feedback is the touch-and-mark feedback, which marks, in real-time, each place on the real image of the breast as soon as it is touched by the fingers. In this way, a woman can clearly understand in real time where have been covered until now, and she can then plan a strategy for the next palpation. The system is now using this type of feedback. The second feedback is to visualize the recorded finger position data graphically in a 3D space. A woman can manipulate it using stereo glasses and spaceball and can view it from arbitrary 3D angles. This will enhance her understanding of breast palpation mechanism and its importance.
(6) Tracking the body movement
During the palpation, a woman will certainly move her body slightly. In order to correlate the finger positions on the breast before and after body movement, the body itself also needs to be tracked. We have selected the nipples as features for tracking the body movement (both displacement and turning) and are implementing this function in the system. Finger positions will be compensated when the body movement is detected.
(7) Low end system
Finally, the system is expected to be widely used by individual women at homes or in schools, clinics, mobile vans etc. by health care providers. The recorded data of palpation may be transmitted to physicians in the hospitals through the connections using telephone lines. These data can then be visualized in the hospitals and evaluated by the physicians. To this end, we plan to implement the system in PCs to make it affordable as well as practical for most women.
Acknowledgment
This research is supported by U.S. Army Grants (DAMD17-94-V-4015, DAMD17-93-3013, and DAMD17-93-3015DAR). The content of this paper does not necessarily reflect the position or policy of the U.S. government. It is also acknowledged that the phantom breast used in this research was provided by the WRS Group, Inc.
References
1. I. Andersson, K. Aspegren, L. Janzon, T. Landberg, K. Lindholm, F. Linell, O. Ljungberg, J. Ranstam, B. Sigfusson, “Mammographic Screening and Mortality from Breast Cancer: the Malmo Mammographic Screening Trial”, BMJ, 297, 943-948, 1988.
2. L. Tabar, G. Fagerberg, H.H. Chen, S.W. Duffy, A. Gad, “Tumour Development, Histology and Grade of Breast Cancers: Prognosis and Progression”, Int. J. Cancer, 66, 413-419, 1996.
3. E.A. Sickles, S.H. Ominsky, R.A. Sollitto, H.B. Galvin, R.L. Monticciolo, “Medical Audit of a Rapid-Throughput Mammography Screening Practice: Methodology and Results of 27,114 Examinations”, Breast Imaging, 175, 323-327, 1990.
4. M. Le Geyte, D. Mant, M.P. Vessey, L. Jones, P. Yudkin, “Breast Self-Examination and the Survival from Breast Cancer”, British Journal of Cancer, 66, 917-918, 1992.
5. D. Mant, M.P. Vessey, A. Neil, K. McPherson, L. Jones, “Breast Self-Examination and Breast Cancer Stage at Diagnosis”, British Journal of Cancer, 55, 207-211, 1987.
6. V.F. Semiglazov, V.M. Moiseenko, “Breast Self-Examination for the Early Detection of Breast Cancer: A USSR/WHO Controlled Trial in Leningrad”, Bulletin of the World Health Organization, 65(3), 391-396, 1987.
7. S.W. Fletcher, M.S. O’Malley, J.L. Earp, T.M. Morgan, S. Lin, D. Degnan, “How Best to Teach Women Breast Self-Examination: A Randomized Controlled Trial”, Annals of Internal Medicine, 112, 772-779, 1990.
8. J. J. Kuch and T. S. Huang, “Vision-Based Hand Modeling And Tracking For Virtual Teleconferencing And Telecollaboration”, Proc. 5th International Conference on Computer Vision, 1995.
9. J. M. Rehg and T. Kanade, “Visual Tracking Of High DOF Articulated Structures: An Application To Human Hand Tracking”, Proc. 3rd European Conference on Computer Vision, 1994.
10. T. Darrell and A. Pentland, “Space-time Gestures”, Proc. IEEE International Conferences on Computer Vision and Pattern Recognition, 1993.
11. J. Davis and M. Shah, “Recognizing Hand Gesture”, Proc. 3rd European Conference on Computer Vision, 1994.
12. F. Perez and C. Koch, “Toward Color Image Segmentation in Analog VLSI: Algorithm and Hardware”, International Journal of Computer Vision, 12(1), 17-42, 1994.
13. Y. Ohta, “Knowledge-based Interpretation of Outdoor Natural Color Scenes”, Research Notes in Artificial Intelligence 4, Pitman Advanced Publishing Program, 1985.
14. M. Barth, S. Parthasarathy and J. Wang, “A Color Vision System for Microelectronics: Application to Oxide Thickness Measurements”, Proc. IEEE International Conference on Robotics and Automation, 1986.
15. J. R. Kender, “Saturation, Hue and Normalized Color: Calculation, Digitization and Use”, Computer Science Technical Report, Carnegie-Mellon University, 1976.
16. G. J. Klinker, “A Physical Approach to Color Image Understanding”, A.K. Peters, Ltd., 1993.
17. S. A. Shafer, T. Kanade, G. J. Klinker and C. L. Novak, “Physics-based Models for Early Vision by Machine”, Proc. SPIE Vol. 1250, Perceiving, Measuring, and Using Color, 222-235, 1990.