Constellation Matching System and Method

1. Technical Field

This invention relates to systems and methods for pattern matching and, more particularly, to a system and method for matching a set of detected points to a set of reference points in real time.

2. Discussion

Pattern, or constellation, matching problems arise in a variety of settings. One example of a pattern matching problem is in radar seeker systems in which detected points in a scene form a known pattern or constellation of points. The problem then is to match a new set of detected points to the known constellation. Similar problems exist with infrared seeker systems and also in systems for recognizing patterns in visual images. In a typical pattern matching problem, a known pattern or set of points are established. An attempt is then made to match a new set of data to the reference pattern. One conventional approach to solving pattern matching problems involves complex scene recognition algorithms. These algorithms are typically very time consuming to develop, often require special processors, and may not be fast enough for real-time application. Other approaches to pattern matching use a human at some point in the loop. This creates problems in speed, reliability, etc.

Thus, it would be desirable to provide a system for constellation matching which can be implemented in real time. In addition, it would be desirable to have a pattern matcher which minimizes the amount of processing necessary and avoids the need for special purpose hardware. This would permit the use of conventional processors already installed in existing systems or vehicles.

Pursuant to the present invention, a system and method is provided for matching a set of detected points to a reference point pattern. This system includes a programmable processor for storing the coordinates of reference points from the reference pattern. These reference coordinates are defined in a reference frame. In addition, the processor stores the coordinates of the detected points. The detected coordinates are defined in a detected coordinate frame. Next, the points in the reference coordinate frame are transformed to the detected coordinate frame using an apriori coordinate transform. Vector tables of each possible ordered pair of reference points and detected points are then created. A list of pairs of matched reference and detected vectors is constructed. The criteria for matching vectors is that the magnitude of the difference vectors between pairs of reference and detected vectors is less than a predetermined threshold.

A match matrix histogram is then created which contains bins for counting each Detection--Reference point correspondence generated by relating the Heads and Tails of the matched Detection--Reference vector pairs. Finally, matches are selected in the matrix by eliminating less probable matches based upon the number of counts in the bins and the geometrical consistency of the implied point matches.

The various advantages of the present invention will become apparent to one skilled in the art, by reading the following specification and by reference to the drawings in which:

FIG. 1 is a block diagram of the overall constellation matching system in accordance with the present invention;

FIG. 2 is a diagram of a set of reference points forming a reference constellation;

FIG. 3 is a diagram of a set of detected points to be matched with the reference points shown in FIG. 2;

FIG. 4 is a first pass of a match matrix histogram utilized by the constellation matcher in FIG. 1;

FIG. 5 is a table of matched vectors in accordance with the present invention;

FIG. 6 is a refined version of the match matrix histogram shown in FIG. 4;

FIG. 7 is a diagram of the computation of median offsets in accordance with the present invention;

FIG. 8 is a diagram of the use of median offsets to check row MAX matches for consistency;

FIG. 9 is a flowchart of the initialization routine in accordance with the present invention;

FIG. 10 is a flowchart of the detection routine in accordance with the present invention;

FIG. 11A-B are flowcharts of the matching routine in accordance with the present invention;

FIG. 12 is a flowchart of the update routine in accordance with the present invention;

FIG. 13A is a drawing of reference and detected points for the real-time example in accordance with the present invention;

FIG. 13B is a drawing of the coordinate system for transforming input frame to REF frame.

FIG. 14 is an input reference point table in accordance with the present invention;

FIG. 15 is a reference point table derived from the input reference point table in FIG. 14;

FIG. 16 is an input reference aimpoint table in accordance with the present invention;

FIG. 17 is a reference aimpoint table derived from the input reference aimpoint table of FIG. 16;

FIGS. 18A and 18B are a series of tables illustrating the Detection Point/Vector/Quantization Coordinate Definition;

FIG. 19 is a reference vector table in accordance with the present invention;

FIG. 20 is a quantization pointer table in accordance with the present invention;

FIG. 21 is a detection points linked list in accordance with the present invention;

FIG. 22 is a DET/REF vector match linked list in accordance with the present invention;

FIG. 23 is a potential match matrix in accordance with the present invention;

FIG. 24 is a refined potential match matrix derived from the potential match matrix in FIG. 23;

FIG. 25 shows refined potential match matrix row statistics and REF/DET offset vectors table in accordance with the present invention;

FIG. 26 shows a final point match table in accordance with the present invention; and

FIG. 27 is an estimated DET frame aimpoint table in accordance with the present invention.

The basic function of the constellation matching system, in accordance with the present invention, is to locate a pattern of reference points within an image of detected points and then to use the correspondence information to locate a specific point or points of interest in the detection image. Referring now to FIG. 1, the overall scheme of the constellation matching system 10 is shown. A reference point storage means 12 contains coordinates for the reference points in a well defined coordinate system. In addition, the coordinates of the specific point of interest are also supplied to the reference point storage means 12. The coordinates of the reference points and the specific point (aimpoint) are transmitted to a constellation matching processor 14.

Overall, the general approach of the present invention is to proceed in four basic steps: Initialization, Detection, Matching and Update. During the Initialization period the constellation matching processor 14 will perform all computations that can be performed prior to the knowledge of the detection points. A sensor 16 will then provide coordinates of detected points (in a detection coordinate system) to the constellation matching processor 14. The sensor 16 may be gathering radar, IR, or other information to generate detected point coordinates.

All computations that can be made based on knowledge of a single detection point, will be made by the constellation matching processor 14 during an ongoing Detection period as time permits. Then all remaining computations based upon the entire detection image will be made during a periodic Matching period as detection points are sequentially added to the detection image. Next, purging of "extinct" detection point data and reinitialization will be performed after each unsuccessful matching attempt during an Update period. Once a final match has been derived by the constellation matching processor 14 the coordinates of the matched points will be used to determine the location of the specific aimpoint in the detected image. The coordinates of the aimpoint may then be transmitted to a control system 18 which may be, for example, a missile guidance system or other means of utilizing the aimpoint information.

While the present invention is useful in many general cases, the preferred embodiment, described below, will be presented for the case where the reference and detected images are two-dimensional and involve only minimal differences in viewing perspective (rotation and warping) or scaling. This a simple but useful model for many real situations.

Referring now to FIG. 2, a simple constellation 20 consisting of four reference points 22 is shown. An aim point 24 is located in a given, fixed, known location with respect to the orientation of the constellation 20. Coordinates for each of the points 22 and as well as the aimpoint 24 are given in a particular reference coordinate system. FIG. 3 shows a set 26 of six detected points 28. The detected points 28 are defined by coordinates in a detected reference frame. An apriori encounter geometry is specified; thus, the angle of bearing of the detected coordinate frame is known with respect to the reference coordinate frame. An appropriate coordinate transformation is established and the supplied reference coordinates are converted to the apriori detection coordinate frame. These points are referred to as the REF coordinates of the reference points. A North, East orientation for X, Y with an arbitrary origin is used. Also, the apriori encounter bearing (angle relative to North) will set the internal X axis orientation.

Next, a set of vectors is established for each possible ordered pair of reference points. Vectors go from one point to another, and each point pair is nominally used twice. In order to improve throughput, "very long" vectors oriented predominantly "backward" may be eliminated since the corresponding detection vectors could not match them anyway due to the sequential manner in which they are formed. For example, REF vectors with a negative X component exceeding a predetermined magnitude need not be retained.

Next, the storage depth (along the X axis) required to encompass the reference image is computed and the "extinction" criteria are established for detection points. Then a potential match threshold is established which will define counts in a potential match histogram based upon the number of reference points supplied and the apriori probability of point detection.

Referring now to FIG. 4, a potential match matrix 30 is constructed with all of its bins set to 0. The rows 32 in the matrix correspond to detected (DET) points and columns 34 in the matrix 30 correspond to REF points. Other miscellaneous constants, tables and variables which require Initialization are established and initialized during this period. Also, data needed for aimpoint estimation which can be pre-computed for all outcomes of the matching process can be made and stored at this time.

As each new detection point is encountered, a new row 32 in the match matrix 30 is selected from a list of available free rows and the row is designated as "in use". A set of vectors are then established for each possible pair of detected points. Vectors go from each of the old "non-extinct" points to the new point. As each new set of DET vectors is generated, a vector match is attempted between each DET vector and each REF vector. A match between two vectors occurs when the length of their difference vector is less than a predetermined threshold. In order to reduce throughput, a preliminary screening of match attempts is accomplished using a look-up procedure. During the Initialization period, the DET vector space is partitioned into a predetermined number of rectangular regions. Then for each such region, a list of REF vectors which could possibly match any DET vector falling in that region is stored in a table. Then during the Detection period, only a small number of potential REF to DET vector matches need be attempted. Additionally, by testing X and Y separately with a slightly looser criteria and then only testing the length for the few potential matches surviving their initial screening, one can substantially reduce the required throughput. A list of current vector matches is augmented with the new matches.

Referring now to FIG. 5, this list 36 of vector matches is shown. To illustrate, the first entry in the list of vector matches indicates that the vector beginning at detected point 1 and ending with detected point 2 in FIG. 3 has been found to match with the vector beginning with reference point 1 and ending with reference point 3 shown in FIG. 2. From visual inspection it can be seen that these vectors are roughly the same length and direction.

Each vector match yields two corresponding point matches thus, the first entry in the vector match list 36 will yield a match between detected point number 2 and reference point number 3 and also between detected point number 1 and reference point number 1. The match matrix 30 in FIG. 4 will accordingly have its bins incremented to tabulate these results. Thus, a 1 appears at the intersection of the row for detected point number 2 and the column for reference point number 3. Also, a 1 is placed in the bin at the intersection of the row for detected point number 1 and the column for reference point number 1. In FIG. 4 the histogram matrix 30 shows the results of tabulating all matches found in the vector match table 36. It can be seen in the match matrix 30 that for certain detection and reference points there were no matches and that for some there were a maximum of three matches.

The match matrix 30 is now scanned to determine the maximum value contained (MAX). As discussed above, during the Initialization period, a potential match threshold was established. For example, as shown in FIG. 4, this threshold is established to be 2.25. Now, the MAX, which is equal to 3, is compared to the match threshold. If this MAX is less then the potential match threshold, then the matching process terminates for this attempt without finding the reference pattern. Otherwise a bin retention threshold is established and the process continues below. As shown in FIG. 4, the MAX is greater than the threshold and the bin retention threshold (REQ) is calculated as half the MAX, or 1.5.

Next, the match matrix 30 is scanned again and any bin count which falls below the established threshold has its corresponding bin in a new refined potential match matrix 38 shown in FIG. 6, set to -1 (all other bins are set to 0). This flags the matrix refinement process to exclude "weak" vector matches which would normally increment this bin. Thus, each of the bins in FIG. 4 having only a 1, do not meet the 1.5 retention threshold and are underlined. Likewise, in the vector match table 36 in FIG. 5, these matches are identified under the column labeled "drop".

The refined potential match matrix 38 is completed by scanning the list of current vector matches and incrementing the two corresponding bins only if neither bin to be incremented is currently negative. This eliminates the underlined bins in the match matrix 30 and may reduce the counts in other bins as well in the refined potential match matrix. It should be noted that the BIN'S do not get reduced directly. They do get reduced indirectly via the refinement process. Since 2 BIN'S are eliminated for each vector match that is eliminated, other BIN'S can be reduced also. Next, the refined potential match matrix 38 is scanned to determine the maximum value (MAX) contained. The maximum value for each row plus the sum of the row entries is also determined for this scan. The column index for the maximum bin in each row is then determined. If more than one entry is at the row maximum value, the column index is set to minus one.

Next, a list of "high confidence" point matches is generated from the revised potential point match matrix 38 based on the first criteria which applies from the selection below:

1. If any row in the matrix 38 has a single non-zero value and that value is equal to the matrix maximum (MAX) then select all such point matches in the matrix. There could be up to one of these per row.

2. If any row in the matrix 38 has only one value equal to the matrix maximum (MAX) then select all such point matches in the matrix. There could be up to one of these per row.

3. Select all point matches in the matrix 38 whose bin value is equal to the matrix maximum (MAX).

The appropriate preliminary REF to DET transformation calculations are performed at this point to prepare for the final point match selection process. Note that for now, a median offset vector (REF to DET) is established from the set derived from the high confidence point matches generated above. X and Y medians are determined separately. It will be appreciated that while the above coordinate transformation converted the reference points to the correct bearing angle of the detected point frame, there still will be an offset between the reference and detected points. For example, FIG. 7 shows that the high confidence matches (2, 2) (3, 3) and (6, 4) represented by the solid circled bins in the refined matrix 38 have their median offsets X, Y computed.

The final point matches are selected by transforming each potential REF match point to the DET frame using the preliminary transformations defined above. That is, the median offset vector is added to REF point coordinates to estimate their expected DET frame values. The corresponding DET match point is checked for spatial consistency (the length of the difference vector lies within a predetermined tolerance). The potential matches to be tested during this final path are those whose non-zero bin values are equal to their row maximums. A list of final point matches and a final match count is then generated. This process is illustrated in FIG. 8.

If the number of final point matches is insufficient, then the process terminates. Otherwise the reference pattern is said to be found and the process proceeds to the aimpoint location process described below.

The appropriate REF to DET transformation calculations are performed using the final point match pairs. A more sophisticated transformation is used at this stage to account for any minor perspective distortions and small differences in scale and rotation. The transformation is then used to locate the specific aimpoint in the DET image.

Detected points which have become extinct (X coordinates less than predetermined minimum) must be eliminated. This is accomplished by decrementing the appropriate bins in the permanent potential match matrix which arose from the extinct point then, by removing the corresponding vectors which contain the extinct point as an end point from the vector match list of current detection points. Further, the rows of the matrix and the point and vector lists should be appropriately compressed. The actual compression may be by chain pointer modification. The vacated rows should contain all zeros.

FIGS. 9 through 12 are flowcharts of the above-described process. In particular, the Initialization period is summarized in the flowchart in FIG. 9. The Detection process is summarized in the flowchart in FIG. 10. The Matching Process is summarized in FIGS. 11a-11b and the Update Process is summarized in the flowchart in FIG. 12.

A simple aimpoint estimation logic example for the Constellation Matcher 14 will now be reviewed. This case assumes unknown translation and negligible rotation or scale factors on either axis. The estimated coordinates of the aimpoint are found by the following transformation, where the centroids for all final matched pairs are computed as follows: ##EQU1##

Note that the above case is simple but is very dependent on assumptions. A more complex but also more tolerant approach will now be reviewed. This is for the case of Unknown Translation, Potentially Very Small Rotation (.ltoreq.5.degree.), No Y Scale Factor, and Potentially Small X Scale Factor (.ltoreq.10). ##EQU2##

It will be appreciated that a more generalized transformation may also be used. For example, a least squares fit estimate of a transformation matrix that includes rotation and scale and shift (as well as perspective) may be employed.

An example of one implementation of the present invention for real-time constellation matching will now be discussed. It will be appreciated that this example is easily implemented in software on conventional computers to achieve rapid solutions to the matching problem. Referring now to FIG. 13A, the example shows five reference points and two aimpoints. FIG. 13B shows the rotation of the coordinates by a bearing angle of 90.degree. after rotation. FIG. 14 indicates the coordinates of the input reference point table upon rotation, the coordinates become the reference point table coordinates indicated in FIG. 15. Likewise, FIG. 16 indicates the reference aimpoint table and FIG. 17, the coordinates of the reference aimpoint table upon rotation by 90.degree..

All definitions for terms used in the real-time implementation example in FIGS. 13-27 inclusive are as follows.

Vectors which have XRVECT<SVCUTOFF need not be retained in the table since they could never match a corresponding DET vector (HEAD=newest DET point; TAIL =Previous DET point)

This table consists of consecutive lists of REF vector indices which could potentially match DET vectors falling in the corresponding quantization BIN. Lists for BINS are stacked contiguously in the table.

A Reference Vector Table is now created which contains all of the ordered pairs of reference points after eliminating those vectors having an XV less than XV cut-off where XV cut-off is equal to -75. In particular eight possible vectors were eliminated and 12 remain as shown in FIG. 19.

In this example, the storage depth for detection points (X) is DETDEPTH=900 for the Y axis DETWIDTH=1000 and the width of a single detect scan is SCANSIZE=75; XQSTEP=512, YQSTEP=512, NQXSTEPS=2 and NQYSTEPS=4. Also VMERRLIM=60, XDVQOFF=75 and YDVQOFF=1000.

FIG. 18 contains the Detection Point/Vector/Quantization coordinate definition tables as well as the constraints used in this example. The quantization pointer table and matchable reference vectors table are shown in FIG. 20. The purpose of quantization is to speed-up throughput. If all possible vector matches were tried, the number of trials would get very large due to the combinatorics. Since there are really very few vectors that are close to matching, the process can be streamlined by using quantization. Thus, all combinations of the reference points are used to make reference vectors and then they are divided up into quantized bins. Then for each quantization bin, a list of indices for all REF vectors which could possible match any DET vector corresponding to that bin is stored in a table. When a detected vector is determined, it is quantized and the system then tries to match only those REF vectors whose indices were stored in the corresponding quantization bin. Thus, the system knows which ones are even close to matching and the system does not have to look at all of them. In essence, the quantization pointer table provides a way of looking things up quickly. If enough memory were reserved to handle all possible matches, it would consume a great deal of memory storage space.

FIG. 21 shows the detection points linked list. FIG. 22 shows the DET/REF vector match linked list which is the set of all vector pairs that matched based on their difference vectors being small enough. FIG. 23 shows the potential match matrix where the bin MAX is 4, the MXTHRESH is 2.0625 and thus the BIN MAX is greater than the MXTHRESH. Also, the BRTHRESH=0.5x4 is equal to 2 thus, bins less than 2 are excluded to yield the refined potential match matrix shown in FIG. 24. FIG. 25 shows the refined potential match matrix row statistics which are used for active rows to determine where conflicts exist. FIG. 25 also shows the offset vector table. Finally, FIG. 26 is the final point match table indicating that five matches were found. To estimate the detection frame aimpoint coordinates in this simple case, the median REF to DET frame offsets are used. This yields the aimpoint coordinates shown in FIG. 27.

It should be recognized that the constellation matching system 10 can be used for matching a set of reference points with a second set of detected points in a wide variety of settings. Also, the present invention can be used in real-time in many applications. While the above description constitutes the preferred embodiments of the present invention, it will be appreciated that the invention is susceptible to modifications, variation and change without departing from the proper scope and fair meaning of the accompanying claims.