`Adaptive gate multifeature Bayesian statistical tracker
`
`W. B. Schaming
`W. B. Scheming
`RCA Advanced Technology Laboratories
`RCA Advanced Technology Laboratories
`Camden, New Jersey 08102
`Camden, New Jersey 08102
`
`Abstract
`Abstract
`
`A statistically based tracking algorithm is described which utilizes a powerful segmenta(cid:173)
`A statistically based tracking algorithm is described which utilizes a powerful segmenta-
`tion algorithm. Multiple features such as intensity, edge magnitude, and spatial frequency
`tion algorithm.
`Multiple features such as intensity, edge magnitude, and spatial frequency
`are combined to form a joint probability distribution to characterize a region containing
`are combined to form a joint probability distribution to characterize a region containing
`a target and its immediate surround. These distributions are integrated over time to pro(cid:173)
`a target and its immediate surround.
`These distributions are integrated over time to pro-
`vide a stable estimate of the target region and background statistics. A Bayesian decision
`vide a stable estimate of the target region and background statistics.
`A Bayesian decision
`rule is implemented using these distributions to classify individual pixels as target or
`rule is implemented using these distributions to classify individual pixels as target or
`nontarget. An adaptive gate process is used to estimate desired changes in the tracking
`nontarget.
`An adaptive gate process is used to estimate desired changes in the tracking
`window size.
`window size.
`
`Introduction
`Introduction
`
`This paper documents progress during the past year toward the development and demonstra(cid:173)
`This paper documents progress during the past year toward the development and demonstra-
`tions of a statistical tracking algorithm. Papers^ '2 presented in 1981 described some of
`tions of a statistical tracking algorithm. Papers ,2 presented in 1981 described some of
`the initial concepts in this development. Since that time, the statistical tracking algo(cid:173)
`the initial concepts in this development.
`Since that time, the statistical tracking algo-
`rithm has been expanded to incorporate (a) the simultaneous use of multiple features, (b) an
`rithm has been expanded to incorporate (a) the simultaneous use of multiple features,
`(b) an
`adaptive gate process for control of the window size, and (c) positional dependence of the
`adaptive gate process for control of the window size, and (c) positional dependence of the
`misclassification cost factor.
`misclassification cost factor.
`
`The tracking algorithm is based on the use of multifeature joint probability density
`The tracking algorithm is based on the use of multifeature joint probability density
`functions for the statistical separation of targets from their background. The features
`functions for the statistical separation of targets from their background.
`The features
`currently being used are intensity, edge magnitude, and a pseudo spatial frequency feature.
`currently being used are intensity, edge magnitude, and a pseudo spatial frequency feature.
`These features are combined to form the joint distributions which characterize a target
`These features are combined to form the joint distributions which characterize a target
`region and its immediate surround. The distributions are integrated over time to provide
`region and its immediate surround.
`The distributions are integrated over time to provide
`a stable estimate of the target and background statistics. A Bayesian decision rule is im(cid:173)
`a stable estimate of the target and background statistics.
`A Bayesian decision rule is im-
`plemented using these distributions to classify individual pixels as target or nontarget
`plemented using these distributions to classify individual pixels as target or nontarget
`within a tracking window. An adaptive gate process is used to estimate desired changes in
`within a tracking window.
`An adaptive gate process is used to estimate desired changes in
`the tracking window size. The algorithm at present assumes manual target designation.
`the tracking window size.
`The algorithm at present assumes manual target designation.
`
`RCA believes this tracking process is capable of operation in all environments; insensi(cid:173)
`RCA believes this tracking process is capable of operation in all environments; insensi-
`tive to target type, signature, and orientation.; applicable to a variety of sensors; and
`tive to target type, signature, and orientation; applicable to a variety of sensors; and
`extendable to multisensor processing and readily implementable.
`extendable to multisensor processing and readily implementable.
`
`Preprocessing and A/D conversion
`Preprocessing and A/D conversion
`
`The video preprocessing function is an important part of any imaging sensor system, but
`The video preprocessing function is an important part of any imaging sensor system, but
`is more critical when the sensor is an IR device which may exhibit very high dynamic range
`is more critical when the sensor is an IR device which may exhibit very high dynamic range
`capability. In this case it is insufficient to perform a simple AGC based upon global
`capability.
`In this case it is insufficient to perform a simple AGC based upon global
`statistics because the subsequent rescaling to reduce the dynamic range will destroy the low
`statistics because the subsequent rescaling to reduce the dynamic range will destroy the low
`contrast local detail. Instead, some form of local adaptive contrast enhancement should be
`contrast local detail.
`Instead, some form of local adaptive contrast enhancement should be
`applied in which the gain varies with the local contrast. Lo^ simulated and compared sever(cid:173)
`applied in which the gain varies with the local contrast.
`Loa simulated and compared sever-
`al such techniques.
`al such techniques.
`
`Although necessary in a hardware implementation, this function has not been included in
`Although necessary
`in a hardware implementation, this function has not been included in
`the simulations reported here. Ten-second image sequences were digitized from video tape
`the simulations reported here.
`Ten -second image sequences were digitized from video tape
`via an analog video disc and an image processing system. The input to the image processing
`via an analog video disc and an image processing system.
`The input to the image processing
`system was passed through a video processing amplifier so that the levels could be properly
`system was passed through a video processing amplifier so that the levels could be properly
`matched to the A/D converter.
`matched to the A/D converter.
`
`Statistical tracking algorithm
`Statistical tracking algorithm
`
`Targets are often separated from their background by a simple thresholding scheme. Some(cid:173)
`Targets are often separated from their background by a simple thresholding scheme.
`Some-
`times the computation of the threshold is quite sophisticated and involves looking at the
`times the computation of the threshold is quite sophisticated and involves looking at the
`statistics of the video signal. However, thresholding is inherently limited in ability as:
`statistics of the video signal.
`However, thresholding is inherently limited in ability as,
`can be seen by the diagrams in Fig. 1. A simple black and white target can be readily
`A simple black and white target can be readily
`can be seen by the diagrams in Fig. 1.
`thresholded to isolate it from its background. On the other hand a gray target cannot be
`On the other hand a gray target cannot be
`thresholded to isolate it from its background.
`thresholded without using a pair of thresholds properly placed to contain the intensity
`thresholded without using a pair of thresholds properly placed to contain the intensity
`levels on the target. This dual threshold in itself is not prohibitive, but rather the prob(cid:173)
`levels on the target. This dual threshold in itself is not prohibitive, but rather the prob-
`lem lies in the ability to place the thresholds at the appropriate levels.
`lem lies in the ability to place the thresholds at the appropriate levels.
`
`68
`68 / SPIE Vol. 359 Applications of Digital Image Processing IV (1982)
`/ SPIE Vol. 359 Applications of Digital Image Processing IV (1982)
`
`Downloaded From: http://proceedings.spiedigitallibrary.org/ on 11/28/2016 Terms of Use: http://spiedigitallibrary.org/ss/termsofuse.aspx
`
`Page 1 of 9
`
`SAMSUNG EXHIBIT 1008
`Samsung v. Image Processing Techs.
`
`
`
`WPM
`
`INTENSITY
`INTENSITY
`
`GRAY
`GRAY
`
`BLACK-,
`
`BLACK
`
`THRESHOLD
`THRESHOLD
`
`WHITE
`WHITE
`
`INTENSITY
`INTENSITY
`
`BLACK
`BLACK
`
`[GRAY
`GRAY
`
`WHITE
`WHITE
`
`(a) THIS TARGET CAN BE EASILY THRESHOLDED
`(a) THIS TARGET CAN BE EASILY THRESHOLDED
`
`(b) THIS TARGET CANNOT BE EASILY THRESHOLDED BUT
`(b) THIS TARGET CANNOT BE EASILY THRESHOLDED BUT
`REQUIRES A PAIR OF THRESHOLDS PROPERLY PLACED.
`REQUIRES A PAIR OF THRESHOLDS PROPERLY PLACED.
`
`SCAN DISTANCE
`SCAN DIS TANCE
`
`SCAN DISTANCE
`SCAN DISTANCE
`
`Fig. 1.
`Fig. 1.
`
`Example showing two postulated targets. One is easily segmented from the back(cid:173)
`One is easily segmented from the back-
`Example showing two postulated targets.
`ground using a single threshold. The other, however, requires two thresholds
`The other, however, requires two thresholds
`ground using a single threshold.
`which are not easily determined. The statistical process provides a separate
`The statistical process provides a separate
`which are not easily determined.
`threshold for each intensity level.
`threshold for each intensity level.
`
`The statistical segmentation process is a technique which provides an improved method for
`The statistical segmentation process is a technique which provides an improved method for
`extracting the target from its background. Figure 2 depicts this process. Shown are two
`Figure 2 depicts this process. Shown are two
`extracting the target from its background.
`histograms, one taken from a window area of the image containing the target and the other
`histograms, one taken from a window area of the image containing the target and the other
`taken from the immediate surround which represents the background. A single feature, in(cid:173)
`A single feature, in-
`taken from the immediate surround which represents the background.
`tensity, is shown in these histograms for illustrative purposes. The shape of the dis(cid:173)
`The shape of the dis-
`tensity, is shown in these histograms for illustrative purposes.
`tribution shown is arbitrary; there are no assumptions made about their actual shape. The
`tribution shown is arbitrary; there are no assumptions made about their actual shape.
`The
`segmentation process makes a separate assessment of each bin in the histogram to determine
`segmentation process makes a separate assessment of each bin in the histogram to determine
`if pixels whose intensity falls in the bin are more likely to be target or background. In
`if pixels whose intensity falls in the bin are more likely to be target or background.
`In
`addition to solving the threshold selection problem, the statistical tracking algorithm pro(cid:173)
`addition to solving the threshold selection problem, the statistical tracking algorithm pro-
`vides a method to both simplify the multimode tracking concept and provide added capability.
`vides a method to both simplify the multimode tracking concept and provide added capability.
`
`The simplification comes about in the following way. State-of-the-art multimode trackers
`The simplification comes about in the following way.
`State -of- the -art multimode trackers
`typically operate a contrast, edge, and correlation tracker in parallel. An executive
`An executive
`typically operate a contrast, edge, and correlation tracker in parallel.
`process may be defined to determine at any given time which tracking mode is providing
`process may be defined to determine at any given time which tracking mode is providing
`the most reliable estimate of target position. The statistical process, as currently de(cid:173)
`The statistical process, as currently de-
`the most reliable estimate of target position.
`fined, eliminates this mode polling process by combining the available features into multi(cid:173)
`fined, eliminates this mode polling process by combining the available features into multi-
`dimensional statistics representing target and background. Consider the use of intensity
`Consider the use of intensity
`dimensional statistics representing target and background.
`and edge magnitude as the two candidate features. In this case the statistical approach
`In this case the statistical approach
`and edge magnitude as the two candidate features.
`encompasses three tracking modes in an integrated single mode without the need to poll the
`encompasses three tracking modes in an integrated single mode without the need to poll the
`performance of the individual processes. When intensity is the best target background
`When intensity is the best target background
`performance of the individual processes.
`separator, the algorithm operates like a contrast tracker. When edge magnitude is pre(cid:173)
`separator, the algorithm operates like a contrast tracker. When edge magnitude is pre-
`dominate it operates similar to an edge centroid tracker. Because the process is searching
`Because the process is searching
`dominate it operates similar to an edge centroid tracker.
`for pixels in the current frame that are statistically similar to those pixels selected as
`for pixels in the current frame that are statistically similar to those pixels selected as
`target in previous frames, the algorithm is in a sense a correlation type process as well.
`target in previous frames, the algorithm is in a sense a correlation type process as well.
`
`The added capability comes from the fact that there are target/background conditions
`The added capability comes from the fact that there are target /background conditions
`which are inseparable using two features independently but are readily separable using the
`which are inseparable using two features independently but are readily separable using the
`In this example,
`same two features jointly. This is illustrated quite simply in Fig. 3.
`This is illustrated quite simply in Fig. 3.
`In this example,
`same two features jointly.
`neither edge magnitude nor intensity can be used independently to separate the target from
`neither edge magnitude nor intensity can be used independently to separate the target from
`background because both flat distributions cover the entire variable range for both features.
`background because both flat distributions cover the entire variable range for both features.
`On the other hand, the joint distribution clearly delineates the two areas.
`On the other hand, the joint distribution clearly delineates the two areas.
`
`69
`SPIE Vol. 359 Applications of Digital Image Processing IV 11982) /
`SPIE Vol. 359 Applications of Digital Image Processing IV (1982) / 69
`
`Downloaded From: http://proceedings.spiedigitallibrary.org/ on 11/28/2016 Terms of Use: http://spiedigitallibrary.org/ss/termsofuse.aspx
`
`SAMSUNG EXHIBIT 1008
`Page 2 of 9
`
`
`
`NUMBER i
`NUMBER
`OF
`OF
`PIXELS
`PIXELS
`
`1
`
`TARGET WINDOW
`TARGET WINDOW
`- HISTOGRAM
`HISTOGRAM
`
`BACKGROUND
`WINDOW
`HISTOGRAM
`
`TARGET
`WINDOW
`
`BACKGROUND
`BACKGROUND
`WINDOW
`WINDOW
`
`Fig. 2.
`Fig. 2.
`
`HISTOGRAM
`HISTOGRAM
`NUMBER OF PIXELS IN EACH INTENSITY GROUP
`NUMBER OF PIXELS IN EACH INTENSITY GROUP
`Example of how histograms are used to separate a target from its background. Each
`Example of how histograms are used to separate a target from its background.
`Each
`bin in the histogram is examined to determine if the intensity value falling within
`bin in the histogram is examined to determine if the intensity value falling within
`that bin are more likely to be target or background. Although this is a single
`that bin are more likely to be target or background.
`Although this is a single
`feature (intensity) example, the same process is used with multiple features in an
`feature (intensity) example, the same process is used with multiple features in an
`N-dimensional histogram representing a joint probability density.
`N- dimensional histogram representing a joint probability density.
`
`INTENSITY
`
`INTENSITY
`INTENSITY
`
`BACKGROUND
`
`BACKGROUND
`
`INTENSITY
`INTENSITY
`
`EDGE MAGNITUDE
`EDGE MAGNITUDE
`
`(A)
`(A)
`
`JOINT DISTRIBUTION OF INTENSITY
`JOINT DISTRIBUTION OF INTENSITY
`AND EDGE MAGNITUDE FOR A
`AND EDGE MAGNITUDE FOR A
`POSTULATED TARGET/BACKGROUND SCENE
`POSTULATED TARGET /BACKGROUND SCENE
`
`(B)
`(B)
`
`INDEPENDENT DISTRIBUTIONS FROM THE
`INDEPENDENT DISTRIBUTIONS FROM THE
`SAME POSTULATED TARGET/BACKGROUND SCENE
`SAME POSTULATED TARGET /BACKGROUND SCENE
`(TARGET AND BACKGROUND DISTRIBUTIONS
`(TARGET AND BACKGROUND DISTRIBUTIONS
`LOOK ALIKE)
`LOOK ALIKE)
`
`Fig. 3. Simple example showing how the use of joint statistics aids in the separation of
`Simple example showing how the use of joint statistics aids in the separation of
`Fig. 3.
`target from background in situations where the use of the features singly fails.
`target from background in situations where the use of the features singly fails.
`
`Figure 4 is a flow diagram of the statistical tracking mode. The preprocessed video is
`Figure 4 is a flow diagram of the statistical tracking mode.
`The preprocessed video is
`used to generate multiple feature images to be used in the decision process. The features
`The features
`used to generate multiple feature images to be used in the decision process.
`are combined into two joint probability density functions for (a) a target tracking window
`are combined into two joint probability density functions for (a) a target tracking window
`and (b) a background window frame. These distributions are the basis of a statistical
`These distributions are the basis of a statistical
`and (b) a background window frame.
`decision process which is used to classify the image pixels inside the tracking window to
`decision process which is used to classify the image pixels inside the tracking window to
`separate the target from the background. In actuality the statistics from previous frames
`In actuality the statistics from previous frames
`separate the target from the background.
`are used in the classification process for the current frame. At the same time, histo(cid:173)
`At the same time, histo-
`are used in the classification process for the current frame.
`grams are generated from the current image frame so that the statistics can be updated for
`grams are generated from the current image frame so that the statistics can be updated for
`processing subsequent frames. At the end of the classification process the segmented image
`At the end of the classification process the segmented image
`processing subsequent frames.
`is analyzed to determine the appropriate error signals as well as the window size and posi(cid:173)
`is analyzed to determine the appropriate error signals as well as the window size and posi-
`tion for the next frame. In parallel with the pixel rate computations for the Nth frame,
`In parallel with the pixel rate computations for the Nth frame,
`tion for the next frame.
`the statistics from the N-lst frame are integrated with past history and a decision rule is
`the statistics from the N-lst frame are integrated with past history and a decision rule is
`generated for the N+lst frame.
`generated for the N+lst frame.
`
`A sample output from the process is shown in Fig. 5. Only two features were used for
`A sample output from the process is shown in Fig. 5.
`Only two features were used for
`this example, namely, intensity and edge magnitude. The total number of bits utilized for
`this example, namely, intensity and edge magnitude.
`The total number of bits utilized for
`the features is seven four for intensity and three for edge magnitude. The edge magni(cid:173)
`the features is seven - four for intensity and three for edge magnitude.
`The edge magni-
`tude used is the absolute value approximation to the Sobel operator.
`tude used is the absolute value approximation to the Sobel operator.
`
`The next few paragraphs describe some of the steps in this process in more detail.
`The next few paragraphs describe some of the steps in this process in more detail.
`
`The first step in the statistical process is the generation of the features to be used.
`The first step in the statistical process is the generation of the features to be used.
`There are many potential candidates, some of which are computationally too burdensome for
`There are many potential candidates, some of which are computationally too burdensome for
`real-time implementation at this time. We therefore have limited our selection of features
`real -time implementation at this time.
`We therefore have limited our selection of features
`
`Computation of features
`Computation of features
`
`70
`/ SPIE Vol 359 Applications of Digital Image Processing IV (1982)
`70 / SPIE Vol. 359 Applications of Digital Image Processing IV (1982)
`
`Downloaded From: http://proceedings.spiedigitallibrary.org/ on 11/28/2016 Terms of Use: http://spiedigitallibrary.org/ss/termsofuse.aspx
`
`SAMSUNG EXHIBIT 1008
`Page 3 of 9
`
`
`
`1
`
`FEATURE
`FEATURE
`
`j1
`
`1
`
`.
`
`.
`
`•
`
`FEATURE
`1
`
`Fl
`FILTER STATISTICS
`NERATE DECISION
`GENERATE DECISION
`GE
`RULE
`LE
`RL
`
`'
`
`1
`UPDATE WINDOW
`UPDATE WINDOW
`POSITION AND SIZE
`POSITION AND SIZE
`AND COSTS FOR NEXT
`AND COSTS FOR NEXT
`FRAME
`FRAME
`
`PREPROCESSEO VIDEO
`PREPROCESSED VIDEO
`
`[
`;
`
`1
`FEATURE
`FEATURE
`2
`
`*
`
`•
`
`i
`
`COMPUTE STATISTICS
`COMPUTE STATISTICS
`CLASSIFY PIXELS
`CLASSIFY PIXELS
`
`1
`
`ANALYSIS OF
`ANALYSIS OF
`SEGMENTED TARGET
`SEGMENTED TARGET
`AREA
`AREA
`
`1
`
`ERROR SIGNALS
`ERROR SIGNALS
`
`Fig. 4. Block diagram of the Bayesian statistical tracking mode. The feature computation,
`The feature computation,
`Block diagram of the Bayesian statistical tracking mode.
`Fig. 4.
`statistics generation, and pixel classification are performed at the pixel rate.
`statistics generation, and pixel classification are performed at the pixel rate.
`The computation of error signals is performed during vertical sync.
`The computation of error signals is performed during vertical sync.
`
`INPUT
`INPUT
`VIDEO
`VIDEO
`
`PREPROCESSING
`PREPROCESSING
`
`• MEDIAN FILTER
`MEDIAN FILTER
`
`TARGET
`TARGET
`SEPARATION
`SEPARATION
`
`• BAYESIAN
`BAYESIAN
`STATISTICAL
`STATISTICAL
`SEGMENTATION
`SEGMENTATION
`
`TRACKING
`TRACKING
`
`• ADAPTIVE GATE
`ADAPTIVE GATE
`• PROJECTIONS
`PROJECTIONS
`
`FEATURE
`FEATURE
`EXTRACTION
`EXTRACTION
`
`• 4 BITS INTENSITY
`4 BITS INTENSITY
`• 3 BITS EDGE
`3 BITS EDGE
`
`Fig. 5. Sample output from the Bayesian statistical tracker simulation using a 64-X-64
`Sample output from the Bayesian statistical tracker simulation using a 64 -x -64
`Fig. 5.
`pixel image of an aircraft at a mountain boundary. Two features were used in the
`Two features were used in the
`pixel image of an aircraft at a mountain boundary.
`statistical segmentation with a total of seven bits.
`statistical segmentation with a total of seven bits.
`
`to those which are readily implemented,
`to those which are readily implemented.
`spatial frequency.
`spatial frequency.
`
`These features are intensity, edge magnitude, and
`These features are intensity, edge magnitude, and
`
`The intensity feature is simply a requantized version of the digitized video signal to
`The intensity feature is simply a requantized version of the digitized video signal to
`obtain the desired number of bits of intensity resolution. The edge magnitude feature is
`The edge magnitude feature is
`obtain the desired number of bits of intensity resolution.
`the sum of absolute values approximation to the Sobel operator. The absolute sum is an
`The absolute sum is an
`the sum of absolute values approximation to the Sobel operator.
`acceptable and computationally more appealing approximation than the true edge magnitude.
`acceptable and computationally more appealing approximation than the true edge magnitude.
`
`The third feature is an approximation to spatial frequency in the horizontal direction.
`The third feature is an approximation to spatial frequency in the horizontal direction.
`Because it is a measure of object size, it could also be considered a simple texture
`Because it is a measure of object size, it could also be considered a simple texture
`measure in a broad sense. The spatial frequency is defined as the function of the run
`The spatial frequency is defined as the function of the run
`measure in a broad sense.
`length where a run is the number of consecutive pixels between which the pixel-to-pixel
`length where a run is the number of consecutive pixels between which the pixel -to -pixel
`difference does not exceed a predefined threshold. The threshold used is the mean value of
`The threshold used is the mean value of
`difference does not exceed a predefined threshold.
`the absolute difference beweeen pixels in the previous frame. The feature value is then
`The feature value is then
`the absolute difference beweeen pixels in the previous frame.
`defined as:
`defined as:
`
`SF = MAXIMUM [0, (2N - RUN LENGTH)]
`SF = MAXIMUM [o, (2N - RUN LENGTH) J
`(1)
`where 2^ is the number of levels into which the spatial frequency feature will be quantized.
`where 2N is the number of levels into which the spatial frequency feature will be quantized.
`
`SPIE Vol 359 Applications of Digital Image Processing /V (1982) /
`SPIE Vol. 359 Applications of Digital Image Processing IV (1982) / 71
`71
`
`Downloaded From: http://proceedings.spiedigitallibrary.org/ on 11/28/2016 Terms of Use: http://spiedigitallibrary.org/ss/termsofuse.aspx
`
`SAMSUNG EXHIBIT 1008
`Page 4 of 9
`
`
`
`An example of the spatial frequency feature is shown in Fig. 6. An arbitrary function
`An example of the spatial frequency feature is shown in Fig. 6.
`An arbitrary function
`is plotted to represent the image intensity I at successive pixels in the x direction.
`is plotted to represent the image intensity I at successive pixels in the x direction.
`Beneath the plotted data are shown the actual pixel intensities, absolute differences, run
`Beneath the plotted data are shown the actual pixel intensities, absolute differences, run
`lengths, and feature values. The threshold used to compute run lengths in the example
`The threshold used to compute run lengths in the example
`lengths, and feature values.
`is 1.3 and the number of quantization levels is 8(N « 3 bits). The first sample-to-sample
`is 1.3 and the number of quantization levels is 8(N = 3 bits).
`The first sample -to- sample
`difference which exceeds the threshold 1.3 is the sixth sample. Samples 1 to 5 represent
`Samples 1 to 5 represent
`difference which exceeds the threshold 1.3 is the sixth sample.
`a run of length 5 in which the differences do not exceed threshold. The corresponding
`The corresponding
`a run of length 5 in which the differences do not exceed threshold.
`feature value is 3 which is assigned to all pixel locations in the run. The higher feature
`The higher feature
`feature value is 3 which is assigned to all pixel locations in the run.
`values indicate smaller distances between gradient values exceeding threshold. Note that
`values indicate smaller distances between gradient values exceeding threshold.
`Note that
`the low amplitude variation between the pixels 6 and 14 do not exceed the threshold and
`the low amplitude variation between the pixels 6 and 14 do not exceed the threshold and
`therefore do not define the boundary of a run. The feature is intended to provide informa(cid:173)
`therefore do not define the boundary of a run.
`The feature is intended to provide informa-
`tion about the size (in the x direction) of areas or patches which have uniform or slowly
`tion about the size (in the x direction) of areas or patches which have uniform or slowly
`varying intensity.
`varying intensity.
`
`Generation and integration of statistics
`Generation and integration of statistics
`
`Histograms from two separate regions in the image must be computed to provide the prob(cid:173)
`Histograms from two separate regions in the image must be computed to provide the prob-
`ability density functions required by the decision rule. The regions from which the histo(cid:173)
`The regions from which the histo-
`ability density functions required by the decision rule.
`grams are generated are shown in Fig. 7. The assumption in the segmentation algorithm is
`The assumption in the segmentation algorithm is
`grams are generated are shown in Fig. 7.
`that the target is absent from the frame region. For both the frame and window regions a
`For both the frame and window regions a
`that the target is absent from the frame region.
`multifeature histogram is defined as
`multifeature histogram is defined as
`
`HFR (fl, f2, f3)
`f 3>
`HFR
`
`Frame Region Histogram
`Frame Region Histogram
`
`HN
`HWR (f
`HWR
`
`2'
`f2, f3)
`
`1,
`for the Nth image in the sequence.
`for the Nth image in the sequence.
`
`Window Region Histogram
`Window Region Histogram
`
`After normalization by the respective areas of the frame and window regions the histo(cid:173)
`After normalization by the respective areas of the frame and window regions the histo-
`grams become the discrete joint probability densities
`grams become the discrete joint probability densities
`
`SENSOR
`SENSOR
`FIELD-OF-VIEW
`FIELD -OF -VIEW
`
`FRAME REGION
`FRAME REGION
`(FR)
`(FR)
`
`WINDOW REGION
`WINDOW REGION
`(WR)
`(WR)
`
`Hp R (fj, f2,13) - multifeature histogram from frame region
`HFR Di, f2, f3) - multifeature histogram from frame region
`
`FWR (f1' f2' ty ~ multifeature histogram from window region
`FAIR (ff, fZ, f3) - multifeature histogram from window region
`
`PFR (f1, f2. f3)
`FR
`
`f2 , f3 ).
`PWR (fl, f2, f3) .
`WR
`
`-
`
`+
`
`X
`
`X
`
`X
`
`X
`
`z
`
`XX
`
`X
`
`X
`
`XTMITY
`zióá3100
`
`V
`
`z
`
`X
`
`,
`
`:
`
`x
`
`X
`
`:
`
`,
`
`7
`
`%
`
`%
`
`%
`
`.
`
`C
`
`X
`
`z
`
`X
`
`-
`
`X
`
`X
`
`X
`
`X
`
`X
`
`X
`
`X
`
`, a
`
`,
`
`,
`
`,
`
`,
`
`',
`
`,
`
`',
`
`,
`
`',
`
`,
`
`,
`
`,
`
`Fig. 6. Sample which shows the procedure
`Sample which shows the procedure
`Fig. 6.
`for calculating the pseudo spatial
`for calculating the pseudo spatial
`frequency feature. The absolute
`The absolute
`frequency feature.
`difference threshold used to compute
`difference threshold used to compute
`run lengths in the example is 1.3,
`run lengths in the example is 1.3,
`which is the average difference. The
`The
`which is the average difference.
`number of quantization levels for
`number of quantization levels for
`the feature is 8.
`the feature is 8.
`
`Fig. 7. Areas of the image over which the
`Areas of the image over which the
`Fig. 7.
`multifeature histograms are com(cid:173)
`multifeature histograms are com-
`puted. It is assumed that the
`It is assumed that the
`puted.
`target is absent from the frame
`target is absent from the frame
`region which is defined as a
`region which is defined as a
`border around the window region
`border around the window region
`containing the target.
`containing the target.
`
`72 / SPIE Vol. 359 Applications of Digital Image Processing IV (1982)
`/ SPIE Vol. 359 Applications of Digital Image Processing IV (1982)
`72
`
`Downloaded From: http://proceedings.spiedigitallibrary.org/ on 11/28/2016 Terms of Use: http://spiedigitallibrary.org/ss/termsofuse.aspx
`
`SAMSUNG EXHIBIT 1008
`Page 5 of 9
`
`
`
`To minimize short-term statistical variations these probability densities are combined
`To minimize short -term statistical variations these probability densities are combined
`in a weighted sum with the past history of the statistics. This fading memory filtering
`in a weighted sum with the past history of the statistics.
`This fading memory filtering
`is performed once each frame so that the statistical updating keeps up with the frame rate
`is performed once each frame so that the statistical updating keeps up with the frame rate
`of the video. The filtering is defined by
`The filtering is defined by
`of the video.
`_
`N
`__
`N-l
`N
`= a PFR + (1 - a) FPN-1
`FPFR
`
`(2)
`
`FPWR
`TA?TD
`
`= b PWR + (1 - b) FPWR1
`TATTD
`'
`~~
`'
`TATTD
`
`(3)
`'
`'
`
`where
`where
`
`Ni?n
`T? <U^WR
`FP , FPTvm are the filtered probability density functions at the Nth frame
`are the filtered probability density functions at the Nth frame
`FPFR, FPWR
`time
`time
`
`N
`N
`P^L' p]!iL
`FR WR
`PFR' PWR
`
`a, b
`a, b
`
`are the unfiltered density functions computed from the current frame
`are the unfiltered density functions computed from the current frame
`N
`N
`
`are the weighting factors which control the amount of smoothing
`are the weighting factors which control the amount of smoothing
`performed.
`performed.
`
`In the simulations performed to date, the filtered statistics up to and including frame
`In the simulations performed to date, the filtered statistics up to and including frame
`N-l are used to generate the decision rule to be used on frame N.
`N -1 are used to generate the decision rule to be used on frame N.
`
`Minimal cost decision rule
`Minimal cost decision rule
`
`The decision rule used in the classification of pixel