Lock-on-After Launch Missile Guidance System Using Three Dimensional Scene Reconstruction

This application claims priority under 35 USC .sctn.119 of provisional application number Ser. No. 60/033,397 filed Dec. 17, 1996.

1. Field of the Invention

This invention relates to a lock-on-after-launch (LOAL) guidance system and, more specifically, to such a guidance system using three dimensional lock-on-after launch (3D.sub.-- LOAL) scene reconstruction.

2. Brief Description of the Prior Art

Much of the prior art relating to LOAL guidance systems relates to variations of the theme of projecting three dimensional (3D) models into their two dimensional (2D) representations and then registering these representations to an image. Such systems are exemplified by Ditzler, W. R., et al., "Multispectral Image Correlation for Air to Ground Targeting," presented at symposium for "Low-Level and Nap of the Earth (N.O.E.) Night Operations, " Rome, Italy, October, 1994.

In the current state of the missile guidance autonomous LOAL algorithms, a painstaking procedure known as "Prebriefing" or "Target Reference Preparation" is required. "Prebriefing" is the act of preparing a description of the target that the algorithm requires in order to locate the target and "prebrief" is the output of prebriefing. In the prebriefing task, the mission planners must follow a rule based approach in which they model a few salient features of the target area which are: (1) unique in shape and (2) visible in contrast to the missile at its preselected lock-on point. Thus, the mission planner must know "geometry" and "signature". Geometry information can be obtained rather easily, but signature further requires identifying the material types of the target objects and factoring in weather (wind, cloud cover, atmosphere, sun angles, etc.) and internal heat sources (heating/cooling, generators, boilers, etc.). Although the main signature question is whether there exists a moderate amount of contrast for certain regions, this can still be extremely time consuming.

Performance level prediction can also be extremely time consuming because extensive Monte Carlo based (randomized) simulations with feedback iteration on the prebriefing must be performed. Once the prebriefing is prepared, the missile is ready for launching.

Although the actual prebriefing model is three dimensional in nature and defined in a world coordinate system, it is only applicable to a limited set of missile approaches and viewing ranges. Once the missile has flown into this "basket" of approach angles and ranges, the three dimensional template is projected into the two dimensional image planes, hidden surfaces are removed and the template is registered to the image. Here, the two dimensional nature of the registration process becomes a limiting factor in the accuracy of aimpoint (the point the missile tries to guide to in the target area) placement. For example, small errors in position or orientation can cause perspective differences in the image that dilute the registration. Fortunately, there is some tolerance to transformation inaccuracies in this method, however this method provides no mathematical mechanism for correcting these transformation errors. For this reason, aimpoints offset from the center of this template must also be kept to short distances because even small rotation errors can result in quite large aimpoint offset errors which degrade overall circular error probable (CEP), the minimum radius around the aimpoint which contains 50 percent of the actual missile impact points. In many cases, multiple templates must be defined for various stages of the mission to overcome these errors and accommodate pixel-on-target constraints due to target size.

In accordance with the present invention, rather than using a two dimensional description of the target (template) to match a two dimensional realization of the target (the sensor image), a three dimensional description of the target area is generated using the sequence of two dimensional sensor images combined with relative movement of the missile. This 3D description is matched to a 3D model of the target area or a 3D description of the target scene is projected into the 2D reconnaissance image and registers in 2D.

In addition, rather than manually modeling only a few unique aspects of the target, the entire target area is automatically modeled. Also, rather than relying upon prediction of a few required contrasts, the invention merely requires location of a sufficient number of object outlines to be unique to the target. Furthermore, rather than having one preset range at which acquisition must take place, acquisition is a continuous process in which the final aimpoint decision can be easily delayed until a high enough level of registration confidence is achieved.

The above is accomplished in conjunction with a 3D.sub.-- LOAL algorithm composed of two parts, one part being 3D.sub.-- Prebriefing (the target area processing that takes place before the missile is launched) and 3D.sub.-- LOAL which occurs after the missile is in flight.

In conjunction with 3D.sub.-- Prebriefing, the algorithm may use several different types of information to acquire aimpoints. These types of information are, in order of desirability, (1) facet based, (2) wire frame based prebrief and (3) recon-photo based prebrief.

In the case of facet based information, individual objects in the target area are described by 3D polygons (facets). Therefore, the algorithm determines which object properties will be visible from any particular viewpoint and eliminates hidden lines from ever being considered in the matching process. This type of information is usually available for "high value" targets. An aimpoint and trackpoint are described as 3D points referenced to this set of 3D facets. The trackpoint is the point on the target the missile sensor is trying to find and the aimpoint is the point it is trying to hit. The geodetic values of each 3D facet for each object are grouped and stored along with the 3D aimpoint as the prebrief file.

In the case of wire frame based prebrief information, 3D endpoints of sides of polygons are available, but object grouping information is not. Therefore, hidden lines cannot be removed. Thus, matching times are longer and discrimination capability is slightly reduced. The aimpoint and trackpoint are described as 3D points referenced to this set of 3D lines. The geodetic values of the 3D line endpoints, the aimpoint and trackpoint are stored as the prebrief file.

In the case of recon-photo based prebrief information, a digitized overhead photo is available, yielding a single 2D description of the target area. First, the image is processed for line features by 2D.sub.-- line.sub.-- finding. Then, a trackpoint and an aimpoint are selected. The extracted lines with a distance D.sup.A of the trackpoint are then calculated. These lines are then highlighted and the aimpoint altitude with respect to the lines is entered into the mission planner. The geodetic values of the 2D line endpoints, the aimpoint and the trackpoint are stored as the prebrief file.

The inflight portion of the algorithm has six processing stages, these stages being (1) 3D.sub.-- LOAL.sub.-- MIDCOURSE, (2) 3D.sub.-- LOAL.sub.-- START, (3) 3D.sub.-- LOAL.sub.-- ACQUIRE.sub.-- 3D.sub.-- LINES, (4) 3D.sub.-- LOAL.sub.-- MATCH.sub.-- PARAMETERS, (5) 3D.sub.-- LOAL.sub.-- LOCK.sub.-- ON and (6) 3D LOAL.sub.-- TRACK.

During the 3D.sub.-- LOAL.sub.-- MIDCOURSE portion (1) of the missile flight, guidance is provided by the navigation hardware of the missile (GPS or inertial). The missile is flown to an area near the target. Once the line-of-sight (LOS) range between the estimated missile and estimated aimpoint falls below the start.sub.-- range.sub.-- value, the 3D.sub.-- LOAL.sub.-- START stage (2) begins.

If the facet-based prebrief.sub.-- model is used, the 3D.sub.-- LOAL algorithm runs in background to remove hidden lines from the prebrief.sub.-- model for the viewing conditions expected at the 3D.sub.-- LOAL.sub.-- START point.

If a recon-based prebrief.sub.-- model is used and only an aimpoint, trackpoint and a raw recon-photo is available, then the 3D.sub.-- LOAL 2D.sub.-- line.sub.-- finding algorithm can run in background and process the recon-photo for 2D.sub.-- lines to be used in the model matching portion of 3D.sub.-- LOAL. This option is included for an operation scenario in which a recon-photo has been datalinked to the missile after its launch.

The 3D.sub.-- LOAL.sub.-- START stage (2) should last for only one frame. Its purpose is to define a 3D reference system in which all 3D line extraction and matching calculations and all tracking propagation calculations take place.

The missile orientation is used to define the terminal coordinate system as follows:

The missile down vector (the vector which emanates from the missile center of gravity and points directly at the center of the earth) is mapped to the terminal coordinate system z vector.

The component of the missile body vector (the vector which emanates from the missile's center of gravity and points through the nose cone of the missile) parallel to the ground plane is mapped into the terminal coordinate system x vector. The cross product of the terminal coordinate system (x.times.z) vectors defines the terminal coordinate system y vector. The origin of the terminal coordinate system is calculated as the preplanned trackpoint.

The terminal coordinate system to the earth centered earth fixed (ECEF) (a right handed three dimensional coordinate system which has its origin at the center of the earth, a Z vector through the north pole and an X vector through the equator/Greenwich meridian intersection) conversion transforms as well as the ECEF coordinate of the terminal coordinate system origin are saved. These values do not change for the remainder of 3D.sub.-- LOAL.

Also, three reference planes (Ph1 and Ph2 for horizontally oriented 3D.sub.-- lines and Pv for vertically oriented lines 3D.sub.-- lines) are calculated. Ph1 and Ph2 are parallel to the terminal coordinate z vector and form 45 degree angles with the x and y directions. Pv is the terminal groundplane. All three planes intersect the terminal system (0,0,0) point.

The purpose of the 3D.sub.-- LOAL.sub.-- ACQUIRE.sub.-- 3D.sub.-- LINES stage is to generate Ng number of well-behaved 3D.sub.-- lines (good.sub.-- 3D.sub.-- lines) for matching against the stored prebrief model or image. Several phases of processing are required to generate these 3D lines. The phases are (1) 2D.sub.-- line.sub.-- finding, (2) 2D.sub.-- line.sub.-- tracking, (3) 3D.sub.-- line.sub.-- creation and (4) 3D.sub.-- line.sub.-- filtering.

In phase (1), for each frame processed, a "2D line.sub.-- finder" must be executed on the incoming sensor image. There are many line-finders in the open literature that can be used, examples being Burns, Brian, et. al., "Extracting Straight Lines," IEEE Trans. on Pattern Analysis and Machine Intelligence, PAMI-8, pp 425-455, July 1986 and Shen, J. and Castan, S., "Further results of DRF Method for Edge Detection", 9th International Conference on Pattern Recognition, Rome, 1988. The line finder used requires the properties of (a) extracting 2D.sub.-- lines to sub-pixel accuracy, (b) generating a defined phase feature, (c) having degree of straightness control and (4) having minimum length control. The set of 2D.sub.-- lines generated on each frame is passed to phase (2) 2D.sub.-- line tracking.

In phase (2), the set of 2D.sub.-- lines from each frame is correlated and tracked over time. To do this, 2D.sub.-- lines from frame N are projected into the terminal groundplane using the missile position and sensor pointing parameters. On frame N+1, these 2D.sub.-- line terminal positions are projected back into the sensor image plane using the frame N+1 missile position and sensor pointing parameters. If a 2D.sub.-- line from frame N+1 matches the position, orientation and length of a projected 2D.sub.-- line from frame N within some tolerance, a correlation occurs. A 2D.sub.-- line that correlates is referred to as a 2D.sub.-- track. Eventually, as 2D.sub.-- tracks are correlated over more and more frames, additional information, namely velocity, can be calculated for the tracks and incorporated into the frame-to-frame propagation. Also, a Kalman filter is applied to the states of the 2D.sub.-- tracks to further refine these states.

The 2D.sub.-- tracks which correlate for the minimum number of frames (N.sub.confirmed) are marked as confirmed.sub.-- 2D.sub.-- tracks. Confirmed.sub.-- 2D-tracks are passed to the next phase where their complete trajectories are recorded for 3D.sub.-- line.sub.-- creation processing. Those 2D.sub.-- lines that do not correlate are used to initialize 2D.sub.-- tracks on the N+1 frame. If these new 2D.sub.-- tracks do not correlate within a small number frames (N.sub.m of n), they are deleted from the 2D.sub.-- track list. Also, 2D.sub.-- tracks that do not correlate over a certain number of frames (N.sub.delete) at any time in the 3D.sub.-- LOAL process are also deleted.

In the phase (3) 3D.sub.-- line.sub.-- creation, the confirmed.sub.-- 2D.sub.-- track states, along with the missile position, orientation and sensor pointing information, are saved on every frame on which a confirmed.sub.-- 2D.sub.-- track update (correlation) occurs. The algorithm continues to buffer the confirmed.sub.-- 2D.sub.-- track states until the missile position has changed enough to generate a sufficient angle for triangulation of the real 3D position of the confirmed.sub.-- 2D.sub.-- tracks. The length of this buffering (N.sub.wait) is customized for each confirmed.sub.-- 2D.sub.-- track and is a function of the missile position at the time the 2D.sub.-- track is confirmed and the expected frame number of the first triangulation snapshot. In general, the value N.sub.wait will be longer for 2D.sub.-- tracks generated at longer ranges. The wait value is generated using a lookup table which is a function of the missile down-range (m.sub.x), cross.sub.-- range (m.sub.y) and altitude (m.sub.z) with respect to the expected aimpoint. A confirmed.sub.-- 2D track that stops updating is deleted after (N.sub.delete.sbsb.--.sub.t) number of frames.

Once a confirmed.sub.-- 2D.sub.-- track has been present for its wait period, it can become a 3D.sub.-- line. Two planes are calculated: Pnow-wait composed of the missile position at frame (N.sub.now-wait) and the two endpoints of the 2D.sub.-- line from frame (N.sub.now-wait) projected into the terminal groundplane; and Pnow composed of the missile.sub.-- position at frame N.sub.now and the two endpoints of the 2D.sub.-- line from frame N.sub.now projected into the terminal groundplane. The intersection of these two planes defines the 3D.sub.-- line. The intersection of the 3D.sub.-- line with the three reference planes is then calculated. The reference plane which generates the largest incidence angle is assigned as the reference for this 3D.sub.-- line. The intercept of the 3D.sub.-- line is defined with respect to this assigned reference plane for the life of the 3D.sub.-- line. At this point, the line state is defined as a slope vector (s.sub.x, s.sub.y, s.sub.z) and an intercept (i.sub.x,i.sub.y, i.sub.z) in the terminal coordinate system. To complete a 3D.sub.-- line definition, two endpoint distances d.sub.1 and d.sub.2 must be calculated. To provide this calculation, first planes P1 and P2, consisting of the missile position from frame N.sub.now, the missile position from frame N.sub.now projected into the terminal groundplane and respectively the 2D.sub.-- line endpoints from frame N.sub.now projected into the terminal groundplane are calculated. Next, the intersection of planes P1 and P2 is found with the line {(s.sub.x,S.sub.y,s.sub.z)(i.sub.x,i.sub.y,i.sub.z)}, yielding the points I1 and I2. Finally, the distances d.sub.1 and d.sub.2 of points I1 and I2 from point (i.sub.x,i.sub.y,i.sub.z) are found. The 3D.sub.-- line is then initialized as the eight state vector ›s.sub.x,s.sub.y,s.sub.z,i.sub.x,i.sub.y,i.sub.z,d.sub.1,d.sub.2 !.

In the phase (4) 3D.sub.-- line.sub.-- filtering, each 3D.sub.-- line is updated on each frame after it is initialized if a history exists for frame (N.sub.now.sbsb.--.sub.now.sbsb.--.sub.wait). An update state is calculated just as before except for the history information used which is newer and only the intercept with the assigned reference plane is required. If a 2D-line endpoint from the new frame is within a buffer area of the edge of the sensor image, an updated distance state calculation is not computed for that endpoint. The new 3D.sub.-- line state is combined with the previous state using a Kalman filter. The variances of the states are then examined to see if they have fallen below a preset threshold (T.sub.qc). If a 3D.sub.-- line passes this quality control check, it becomes a good.sub.-- 3D.sub.-- line. Once a preset number (N.sub.g) of good.sub.-- 3D.sub.-- lines is obtained, the 3D.sub.-- LOAL algorithm transitions to the 3D.sub.-- LOAL.sub.-- MATCH.sub.-- PARAMETERS stage. A 3D.sub.-- line which stops updating before becoming a good.sub.-- 3D.sub.-- line is deleted after (N.sub.delete.sbsb.--.sub.t) number of frames.

In the 3D.sub.-- LOAL.sub.-- MATCH.sub.-- PARAMETERS stage, all the processing described in the prior stage continues (i.e., 3D.sub.-- LOAL continues to update 3D.sub.-- lines and find new ones). In addition 3D.sub.-- LOAL begins the process of matching the set of good.sub.-- 3D.sub.-- lines to a subset of the prebrief 3D.sub.-- model.sub.-- lines. This is accomplished in four phases denoted as (1) pertubate.sub.-- model.sub.-- transform, (2) parse.sub.-- model.sub.-- lines, (3) find.sub.-- line.sub.-- correspondences and (4) angle.sub.-- error.sub.-- processing.

In the pertubate.sub.-- model.sub.-- transform phase (1), because the exact orientation of the missile may not be known at the time the terminal coordinate system is defined, there will be rotation errors between the prebriefed model as it is transformed into the terminal coordinate system and the true orientation of the target area. These errors should be limited to .+-. a few fractions of a degree. The 3D.sub.-- LOAL algorithm generates pertubations to the ECEF.sub.-- to.sub.-- terminal direction cosine matrix (DCM) (a 3.times.3 matrix used to transform 3D vectors from one three dimensional coordinate system to another) matrix on each frame that covers the range of rotation transform errors expected in the navigation and prebriefing systems due to pitch and yaw. These pertubations are applied in a systematic manner to the model transformation so that all regions of the rotation uncertainty area are covered in a timely fashion. An acceptable search strategy for uniform pitch and yaw errors of .+-..beta.p,.beta.y degrees respectively which requires 18 frames is:

do for .alpha.p=(-.beta.p,0,.beta.p)

do for .alpha.y=(-.beta.y,0,.beta.y)

pertubate DCM.sub.ecef.sbsb.--.sub.to.sbsb.--.sub.terminal by (.alpha.p,.alpha.y).

The purpose of the pertubation is to align lines near the edge of the parsed prebrief. If the angle errors are sufficiently small, this step may be eliminated completely since phase (4) (angle.sub.-- error.sub.-- processing) will still generate acceptable results. Otherwise, some combinations may be excluded or rearranged to shorten the timeline. For example, if 2D correspondence (see below) is to be used, then only .alpha.y needs to be perturbed (this is the case when the recon-photo prebrief option is used).

In the parse.sub.-- model.sub.-- lines phase (2), the 3D.sub.-- LOAL algorithm must extract those lines in the 3D prebrief.sub.-- model which correspond to the area covered by the seeker footprint. To do this, all model lines are first transformed into the terminal coordinate system using the pertubated transform. Next, the sensor footprint on the current frame is projected into the terminal groundplane. All model lines are found whose 2D projections into the terminal groundplane (2D.sub.-- model.sub.-- lines) are all or partially contained in the sensor footprint. For this frame, the model.sub.-- lines are marked as parsed.sub.-- model.sub.-- lines.

The find.sub.-- line.sub.-- correspondences phase (3) is divided into a find.sub.-- 2D.sub.-- line.sub.-- correspondence portion and a find.sub.-- 3D.sub.-- line.sub.-- correspondence portion.

In the find.sub.-- 2D.sub.-- line.sub.-- correspondence portion, each good.sub.-- 3D.sub.-- line is projected into the terminal groundplane. For each projected good.sub.-- 3D.sub.-- line, all projected parsed.sub.-- model.sub.-- lines are found which have the same angle .+-. some angle tolerance T.sub..alpha.. For each pair of lines produced by the angle nearness test, a position nearness test is applied. A 2D search box is generated around each projected good.sub.-- 3D.sub.-- line by extending its endpoints and its sides by the position uncertainty of the missile navigation and targeting error. All projected parsed.sub.-- model.sub.-- lines that intersect any part of this region are found. The set of all such pairs of projected good.sub.-- 3D.sub.-- lines matching projected parsed.sub.-- model.sub.-- lines via the described angle and distances test constitute the 2D.sub.-- line.sub.-- correspondences for this frame.

In the find.sub.-- 3D.sub.-- line.sub.-- correspondence portion, for each good 3D.sub.-- line, all parsed.sub.-- model.sub.-- lines are found which have the same angle in 3D space .+-. some angle tolerance T.phi.. For each pair of lines produced by the angle nearness test, a position nearness test is applied. A 3D search volume is generated around each projected good.sub.-- 3D.sub.-- line by extending its endpoints and its sides by the position uncertainty of the missile navigation and targeting error. All parsed.sub.-- model.sub.-- lines that intersect any part of this region are found. The set of all such pairs of good.sub.-- 3D.sub.-- lines matching parsed.sub.-- model.sub.-- lines via the described angle and distances test constitute the 3D.sub.-- line.sub.-- correspondences for this frame.

Phase (4) angle.sub.-- error.sub.-- processing is broken into 2D.sub.-- angle.sub.-- error.sub.-- processing and 3D.sub.-- angle.sub.-- error.sub.-- processing.

In 2D.sub.-- angle.sub.-- error.sub.-- processing, only the yaw angle is pertubated. For each pertubation, the angle differences between the 2D.sub.-- line.sub.-- correspondence pairs are calculated and accumulated in a yaw.sub.-- angle.sub.-- error.sub.-- histogram (the pertubation error must be subtracted from this difference to avoid a bias). The end of the pertubation sequence marks the finish for 3D.sub.-- LOAL.sub.-- MATCH.sub.-- PARAMETERS. On the last frame, the yaw.sub.-- angle.sub.-- error.sub.-- histogram is smoothed and its peak value is estimated. The angle error bin for which this peak occurs is the estimated.sub.-- yaw.sub.-- error between the model and the extracted scene and is applied to all subsequent model transformations in the 3D.sub.-- LOAL.sub.-- LOCK.sub.-- ON stage.

In 3D.sub.-- angle.sub.-- error.sub.-- processing, the procedure is very similar to the 2D version except that the angle difference between 3D.sub.-- line.sub.-- correspondence pairs is decomposed into a yaw, pitch and roll component and accumulated in a yaw.sub.-- angle.sub.-- error.sub.-- histogram, pitch.sub.-- angle.sub.-- error.sub.-- histogram and roll.sub.-- angle.sub.-- error.sub.-- histogram respectively. In this case, three sets of projections must be made and correspondence angle errors accumulated, one for each axis of rotation. Again, the end of the pertubation sequence marks the finish of 3D.sub.-- LOAL.sub.-- MATCH.sub.-- PARAMETERS. The two histograms are smoothed and their peak values become the estimated.sub.-- yaw.sub.-- error and estimated.sub.-- pitch.sub.-- error between the model and the extracted scene. Finally, these estimates are applied to all subsequent model transformations in the 3D.sub.-- LOAL.sub.-- LOCK.sub.-- ON stage.

The 3D.sub.-- LOAL.sub.-- LOCK.sub.-- ON stage is broken into a 2D.sub.-- vector.sub.-- registration portion and a 21/2D.sub.-- vector.sub.-- registration portion.

In the 2D.sub.-- vector.sub.-- registration portion, the 2D.sub.-- line.sub.-- correspondence pairs are operated upon and only calculate a 2D trackpoint.sub.-- correction.sub.-- vector or groundplane xy offset. For each 2D.sub.-- line.sub.-- correspondence pair: (a) the 2D vector that is perpendicular to the projected parsed.sub.-- model.sub.-- line and intersects the midpoint of the projected good.sub.-- 3D-line is determined and (b) the 2D perpendicular bisector of this 2D vector is found. The intersection of all 2D perpendicular bisectors with every other 2D perpendicular bisector is calculated. These intersection are accumulated in a 2D array, where the location in the array indicates the 2D location of the intersection. Next, this array is mean filtered. The location of the maximum value in the mean filtered array times 2 is the 2D trackpoint.sub.-- correction.sub.-- vector estimate.

In the 21/2D vector registration portion, the 21/2D.sub.-- vector.sub.-- registration algorithm adds the z component to the trackpoint.sub.-- correction.sub.-- vector. The 2D bisector intersection array generated above is then searched for N.sub.peaks local maximas. For each of the peaks, all 3D.sub.-- line.sub.-- correspondences are calculated using a much smaller uncertainty value (-10 meters) and the xy offset corresponding to that peak location. For each "non-vertical" good.sub.-- 3D.sub.-- line, a vertical offset vector to the center of the corresponding parsed.sub.-- model.sub.-- line is computed. These vertical values are accumulated in a z.sub.-- vector.sub.-- histogram which is an array sorted by the Z vertical value containing the number of occurrences of each vertical value. There are N.sub.peaks z.sub.-- vector.sub.-- histograms. All of the z.sub.-- vector.sub.-- histograms are mean filtered and the maximum value is found. The bin containing the maximum value is the z component of trackpoint.sub.-- correction.sub.-- vector. The (x,y) value generating that histogram completes the 3D trackpoint.sub.-- correction.sub.-- vector.

When the maximum z value reaches some minimum value (T.sub.lock-on) 3D.sub.-- LOAL transitions to the 3D.sub.-- LOAL.sub.-- TRACK stage. The following actions are involved in this transition:

(a) the trackpoint.sub.-- correction.sub.-- vector is added to the missile trackpoint location.

(b) this new aimpoint is copied to the sensor staring point.

(c) the trackpoint.sub.-- correction.sub.-- vector is mapped into the translation portion of the prebrief.sub.-- model transformation (if the correction to the sensor pointing vector is large, the sensor pointing transition may be made gradually to avoid a sudden loss of matching lines and therefore loss of updates to the trackpoint.sub.-- correction.sub.-- vector).

(d) the target location error uncertainty is collapsed to the missile inertial navigation uncertainty for small distances.

(e) since the missile may not be flying at the trackpoint (trackpoint and aimpoint) the difference between the prebriefed trackpoint and aimpoint is added to the corrected trackpoint to become the new aimpoint at which the missile will guide to.

In the 3D.sub.-- LOAL.sub.-- TRACK stage, the 3D.sub.-- LOAL algorithm continues to execute all portions of the algorithm required to generate good-3D-lines and an aimpoint.sub.-- correction.sub.-- vector. It may, as an options also continue to process updates to the estimated.sub.-- yaw.sub.-- error and/or estimated.sub.-- pitch.sub.-- error. In reality, 3D.sub.-- LOAL.sub.-- TRACK is 3D.sub.-- LOAL.sub.-- LOCK.sub.-- ON at a very high frame rate. The higher frame rate is achieved by two factors, (1) the navigation and targeting errors are greatly reduced by the 3D.sub.-- LOAL.sub.-- LOCK.sub.-- ON snapshots(s) and (2) the shrinking sensor footprint results in fewer model.sub.-- lines being parsed. The only other difference is that the set of parameters, i.e., N.sub.wait, N.sub.delete, Ng, is modified. This phase continues until a blind range is reached at which time the missile finishes its flight via inertial navigation.

It should be understood that though the invention is described with references to direction of a missile, the invention can also be used for passive navigation of aircraft, realtime extraction of 3D site models from reconnaissance image sequences, pattern recognition, robot vision, etc.

FIG. 1 is a schematic diagram representing the two main parts of 3DLOAL system;

FIG. 2 is a schematic diagram illustrating the five main processing phases of the 3DLOAL portion of the system;

FIG. 3 is top and side views of the desired missile trajectory once 3DLOAL starts and as related to the five processing phases;

FIG. 4 is a diagram of the terminal coordinate system and its reference planes as established on the 3D.sub.-- LOAL.sub.-- START frame;

FIG. 5 is a schematic diagram of the 3D.sub.-- LOAL.sub.-- ACQUIRE 3D.sub.-- LINES processing phase;

FIG. 6 is a schematic diagram showing procedures required to generate 3D.sub.-- lines from a frame of sensor imagery;

FIG. 7(A) shows a sample sensor image with parts (B), (C) and (D) showing the resulting images and lines after the edge extraction, edge thresholding and line finding procedures have executed;

FIG. 8 illustrates the calculation of the phase angle for a single line;

FIG. 9 illustrates a 4 quadrant phase feature implementation;

FIG. 10 illustrates the line merging calculation, in part (A) the line extended a preset distance and in part (B) a mergable line is found;

FIG. 11 is a schematic diagram showing the nine procedures required to create 2D.sub.-- tracks from 2D.sub.-- lines;

FIG. 12 illustrates the calculation of updated missile positions and sensor pointing angles in the terminal coordinate system;

FIG. 13 illustrates how the change due to missile position and sensor pointing is used to propagate a 2D.sub.-- track;

FIG. 14 is an illustration of the 2D-track/2D.sub.-- line merging criteria with (A) showing the normal calculation between the 2D.sub.-- track and 2D.sub.-- line, (B) showing the distance calculation between the two and (C) showing the percent overlap calculation;

FIG. 15 is a schematic diagram showing the seven processing procedures required to turn 2D.sub.-- tracks into 3D.sub.-- lines;

FIG. 16 illustrates the change in prespective test used to determine when a 3D.sub.-- line can be initialized;

FIGS. 17(A) and (B) illustrate the calculation of the init plane for a 3D.sub.-- line using the missile position and the projection of th eline endpoints into the groundplane;

FIG. 18 is a schematic diagram of the three sub-procedures required to filter 3D.sub.-- lines so that they become good.sub.-- 3D.sub.-- lines and can be used in the model matching phase of 3D.sub.-- LOAL;

FIG. 19 is a schematic diagram of the 3D.sub.-- LOAL.sub.-- LOCK.sub.-- ON phase where 3D.sub.-- lines from the scene are matched to lines in the prebrief and a missile guidance correction is first generated;

FIG. 20 illustrates the calculation of the sensor footprint parsing area with (A) being the direct sensor footprint projection and (B) being the footprint extension by the targeting uncertainty;

FIG. 21 is a schematic diagram showing the three sub-procedures required to generate the cross-range and down-range components of the aimpoint/trackpoint correction vector;

FIGS. 22(A) and (B) show the calculation of 2D.sub.-- line.sub.-- correspondence distance test with (A) showing the buffer box around the good.sub.-- 3D.sub.-- line and (B) showing the intercept calculation with the parse.sub.-- model.sub.-- lines;

FIG. 23 shows the 2D perpendicular bisector calculation for a 2D.sub.-- line.sub.-- correspondence pair with (A) being the calculation of the normal between the two lines and (B) being the calculation of the perpendicular bisector of the normal;

FIG. 24 shows an example 2D registration score surface as generated by simulation;

FIG. 25 is a schematic diagram of the three sub-procedures required to calculate a 2.5D registration or the altitude correction for the amipoint/trackpoint vector;

FIGS. 26(A) and (B) show the calculation of the 2.5D.sub.-- line.sub.-- correspondence distance test with (A) showing the buffer box around the good.sub.-- 3D.sub.-- line and (B) showing the intercept calculation with the parse.sub.-- model.sub.-- lines; and

FIG. 27 illustrates the calculation of the z offset component for a 2.5D.sub.-- line.sub.-- correspondence pair.

Initially, a three dimensional data base is developed of the target area including the target. This data base is developed, for example, by prior reconnaissance photographs of the target area from as many directions as possible. In this way, a 2D picture of the target area is formulated from any direction of approach to the target. This data base is provided on the missile initially with target location being provided in conjunction with the data base as well as with standard GPS and/or inertial navigation techniques. This is described hereinbelow as "prebriefing".

Since the general location of the target area is known, the missile can be directed into the general area of the target by appropriate programming of the missile navigation system in conjunction with GPS and/or inertial navigation in standard manner before or after firing the missile, the preferred embodiment herein providing such target area location information prior to firing. This is also a part of the prebriefing step.

To explain the invention in conjunction with prebriefing option (1) as set forth above, a 3D wire frame model of the target site is initially generated. The wire frame model of the target area including the target is provided and obtained prior to missile launch by prior air reconnaissance or from other sources and stored in a data base, preferably in the missile. The data generated during the 3D LOAL step of FIG. 1 will then be compared with this data base for 3D matches to ultimately lock onto the target when the match meets some minimum predetermined degree of reliability.

In the 3D LOAL operation, as shown in some detail with reference to FIGS. 2, 5, 6, 11, 15, 18, 19, 21 and 25, upon launch of the missile and during the 3D LOAL midcourse stage, missile flight guidance to a preset location in the region of the target is provided by the missile navigation hardware which is controlled in response to the global positioning system (GPS) or inertially controlled. In the initial flight path of the missile to an area near the target, the missile sensor is commencing staring at the estimated target aimpoint/trackpoint as shown in FIG. 3 where the solid line indicates missile direction at the start of 3DLOAL and the dashed line indicates eventual missile flight path. In this description, the targeting aimpoint and trackpoint are assumed to be the same, however an offset aimpoint may be used by simply adding the delta between the prebriefed trackpoint and prebriefed aimpoint prior to passing it to the missile guidance and control. This allows bland targets to also be prosecuted.

With reference to FIG. 4, in the terminal area setup stage, which should last for only one frame of operation, a local 3D reference system is defined in which all 3D line extraction and matching calculations and all tracking propagation calculations take place. The missile orientation at the time of obtaining the frame, such as by camera or other scene reproducing means, is used to define the terminal coordinate system. Also, three reference planes are defined for calculating line intercepts, these reference planes being preferably, but not limited to, a standard three coordinate system with the coordinates mutually perpendicular to each other.

With reference to FIG. 5, in the 3D line acquisition stage, and N.sub.g number is generated for well-behaved 3D.sub.-- lines (good.sub.-- 3D.sub.-- lines) for matching against the stored prebrief model or image. Four phases of processing are required to generate these 3D lines, these phases being (1) 2D line finding, (2) 2D line tracking, (3) 3D line creation and (4) 3D line filtering. With reference to FIGS. 7 to 10, in the 2D line finding phase, for each frame processed, a 2D.sub.-- line.sub.-- finder must be executed on the incoming sensor image. There are many known line-finders that can be used, examples being Burns, Brian, et. al., "Extracting Straight Lines," IEEE Trans. on Pattern Analysis and Machine Intelligence, PAMI-8, pp 425-455, July 1986 and Shen, J. and Castan, S., "Further results of DRF Method for Edge Detection," 9th International Conference on Pattern Recognition, Rome, 1988. The set of 2D.sub.-- lines generated on each frame is passed to the next phase, which is 2D line tracking. With reference to FIGS. 11 to 14, in the 2D line tracking phase, the set of 2D.sub.-- lines from each frame must be correlated and tracked over time, as, for example, by the algorithm described by Giles, Brent, Miller, Keith, Newton, Scott, "Advanced Tracker Algorithm Description", May, 1995. To do this, 2D.sub.-- lines from frame N are propagated into the sensor image plane using frame N+1 missile position and sensor pointing parameters. If a 2D.sub.-- line from frame N+1 matches the position, orientation and length of a projected 2D.sub.-- line from frame N within some tolerance, a correlation occurs, forming a 2D.sub.-- track. A Kalman filter is applied to the states of a 2D.sub.-- track to further refine such tracks. A 2D.sub.-- track which correlates for a minimum number of frames is marked as confirmed and passed to the next phase. With reference to FIGS. 15 to 17, the next phase is 3D line creation wherein the states of the confirmed.sub.-- 2D.sub.-- tracks, along with missile position, orientation and sensor pointing information, are saved on every frame on which a confirmed.sub.-- 2D.sub.-- track update (correlation). occurs. The algorithm continues to buffer the confirmed.sub.-- 2D.sub.-- track states until the missile position has changed sufficiently to generate a sufficient angle for triangulation of the real 3D position of the confirmed.sub.-- 2D.sub.-- tracks. At this point, two planes are calculated to define a 3D.sub.-- line, P.sub.now, composed of the missile position at frame N.sub.now and the two endpoints of the 2D.sub.-- line from frame N.sub.now projected into the terminal groundplane and P.sub.now-wait composed of the missile position at frame N.sub.now-wait and the two endpoints of the 2D.sub.-- line from frame N.sub.now-wait projected into the terminal groundplane. The intersection of these two planes defines the 3D.sub.-- line. The best intersection of this 3D.sub.-- line with the three reference planes defines its origin. With reference to FIG. 18, in the 3D line filtering phase, each 3D.sub.-- line can be updated on each frame after it is initialized if a history exists for frame N.sub.new.sbsb.--.sub.now-wait An update state is calculated except for the history information used which is newer and only the intercept with the assigned reference plane is required. The new 3D.sub.-- line state is combined with the previous state using a Kalman filter. The variances of the states are then examined to see if they have fallen below a preset threshold (T.sub.qc). If a 3D.sub.-- line passes this quality control check, it becomes a good.sub.-- 3D.sub.-- line.

In the rotation error correction stage, all processing described in conjunction with the 3D line acquisition stage continues (i.e.,3DLOAL continues to update 3D.sub.-- lines and find new lines). In addition, 3DLOAL begins the process of matching (or resolving transformation errors between) the set of good.sub.-- 3D.sub.-- lines and a subset of the prebrief 3D.sub.-- model.sub.-- lines. This stage has four processing phases which are (1) pertubate model transformation, (2) parse prebrief model lines, (3) find line correspondences and (4) process angle error histograms. With reference to the first phase, because the exact orientation of the missile may not be known at the time the terminal coordinate system is defined, there will be rotation errors between the prebriefed model as it is transformed into the terminal coordinate system and the true orientation of the target area. The 3DLOAL algorithm generates perturbations to the ECEF.sub.-- to.sub.-- terminal direction cosine matrix on each frame during this stage of operation. The sequence of perturbations should cover a 99 percent confidence bound for the expected rotation transform errors. With reference to FIGS. 20 and 21, in the parse prebrief model lines phase, the 3DLOAL algorithm must extract those lines in the 3D prebrief.sub.-- model which correspond to the area covered by the seeker footprint. To do this, all model lines are first transformed into the terminal coordinate system using the pertubated transform. Next, the sensor footprint on the current frame is projected into the terminal groundplane. All model lines whose 2D projections into the terminal groundplane (2.sub.-- D.sub.-- model.sub.-- lines) are all or partially contained in the sensor footprint are found and, for this frame, these model.sub.-- lines are marked as parsed.sub.-- model.sub.-- lines. With reference to FIGS. 21 and 22, in the phase of finding line correspondences, both 2D and 2D line correspondences are calculated to efficiently achieve an overall model to scene match. To find 2D correspondences, each good.sub.-- 3D.sub.-- line and each parsed.sub.-- model.sub.-- line are projected into the terminal groundplane. For each projected good.sub.-- 3D.sub.-- line, all projected parsed.sub.-- model.sub.-- lines which have the same angle .+-. some angle tolerance T.sub..alpha. are found. Next, a 2D search box is generated around each projected good.sub.-- 3D.sub.-- line by extending its endpoints and its sides by the position uncertainty of the missile navigation and targeting error. All same angle projected parsed.sub.-- model.sub.-- lines that intersect any part of this region are found. The set of all such pairs of projected good.sub.-- 3D.sub.-- lines matching projected parsed.sub.-- model.sub.-- lines constitute the 2D.sub.-- line.sub.-- correspondences for this frame. Similarly, to find 3D.sub.-- line.sub.-- correspondences, a two angle test must be passed and a search volume must be examined. In the phase of processing angle error histogram(s), according to its orientation, each 3D.sub.-- line.sub.-- correspondence can generate an angle difference measure for two of the three angle error histograms (row, pitch and yaw). Therefore, these angle error histograms are accumulated over the transformation perturbation sequence. At the end of the perturbation sequence, which marks the end of this stage, the histogram(s) are smoothed and their peak(s) are found. These peaks then become the estimate.sub.-- roll.sub.-- error, estimated.sub.-- pitch.sub.-- error and estimated.sub.-- yaw.sub.-- error between the model and the extracted scene. These estimates are applied to all subsequent model transformation in the lock-on stage.

With reference to FIGS. 21 to 27, in the translation error correction or lock-on stage, once the rotation portion of the model transformation error has been eliminated, the translation portion of the error can be found. To accomplish this, 3DLOAL performs a 2 stage vector registration. First, a 2D vector registration operates on the 2D.sub.-- line.sub.-- correspondence pairs and only calculates a 2D aimpoint.sub.-- correction.sub.-- vector or groundplane xy offset. Such an algorithm is set forth by Lundgren, Jim, "Registration Algorithm", March, 1993. Then, a 21/2D vector registration algorithm adds the z component to the aimpoint.sub.-- correction.sub.-- vector. Only the top peaks generated by the 2D vector registration procedure are processed. When the value of the registration reaches some minimum value, 3DLOAL transitions to the terminal tracking stage. The following actions are involved in transition to track: (1) the aimpoint.sub.-- correction.sub.-- vector is added to the missile aimpoint location, (2) this new aimpoint is copied to the sensor staring point, (3) the aimpoint.sub.-- correction.sub.-- vector is mapped into the translation portion of the prebrief.sub.-- model transformation and (4) the absolute location uncertainty is collapsed to the missile navigation drift plus sensor pointing error uncertainty.

In the terminal tracking stage, the 3DLOAL algorithm continues to execute all portions of the algorithm required to generate good.sub.-- 3D.sub.-- lines and an aimpoint.sub.-- correction.sub.-- vector. It may, as an option, also continue to process updates to the estimated.sub.-- yaw.sub.-- error and/or estimated.sub.-- pitch.sub.-- error. In reality, terminal tracking is lock-on at a very high frame rate. The higher frame rate is achieved by two factors, these being (1) the navigation and targeting errors (and therefore correspondence regions) are greatly reduced a the lock on snapshot(s) and (2) the shrinking sensor footprint results in fewer model-lines being parsed. The only other difference is the set of parameters, i.e., N.sub.wait, N.sub.delete, N.sub.g are modified. This phase continues until a blind range is reached, at which time the missile finishes its flight via inertial navigation.

The above described procedure was implemented in a computer simulation and tested against a high resolution synthetic database representing a complex urban scene with an average aimpoint error of less than two meters.

The Appendix attached hereto is a 3DLOAL algorithm in accordance with the present invention which is the actual preferred embodiment.

Though the invention has been described with reference to a specific preferred embodiment, many variations and modifications thereof will immediately become apparent to those skilled in the art. It is therefore the intention that the appended claims be interpreted as broadly as possible in view of the prior art to include all such variations and modifications. ##SPC1##