JPH0831134B2 - Optical correlation system and method - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION "Industrial Application" The present invention relates generally to robot vision and optical correlation systems that use matched filters to identify objects and extract appearance information such as position and size of targets. In particular, it relates primarily to robot vision systems that utilize parallel optical processing.

"Prior Art" The research and application of robot vision systems is accelerating in both commercial and military fields. Today's commercial applications have been identified with respect to machine operation and production lines. In the prior art, robotic manipulators have been utilized in relatively large scale manufacturing assembly lines that perform simple tasks such as machine loading and unloading, component stacking, painting and spot welding. These machines
Generally, it operates with a very specific program instruction set, and the positioning and orientation of the workpiece to be handled can be known quite accurately. The programming of such manipulators was usually relatively complex and the program was only useful if the worm's positioning was kept within relatively precise tolerances.

In recent years, by adding various sensory abilities,
Attempts have been made to increase the adaptability of such manipulators. Tactile and auditory abilities are currently being developed in addition to visual abilities as per the present invention. Distance measurement technology and structured light / binocular vision technology (stru
ctured light and binocular vision techniques)
It has been used in robot vision systems as described above. However, neither of these systems is particularly useful in applications for recognizing a target and determining its position and orientation. Furthermore, known robot vision systems require significant processing time from visual detection to confirmation of the presence of a target.

In the prior art, the ability of many robots or machine vision systems to perform their specific directed tasks is disclosed and discussed. These systems are in fact a hybrid of sensors, usually analog means, with mostly digital processing and joint control. A further division of this technique is the choice of whether the object recognition is digital or analog.
Digital systems often rely on video inputs and algorithms to classify objects and parts by size and appearance. Memory libraries are limited only by the amount of computer memory. Fewer optical systems are available, but most rely on some electronic processing.

U.S. Pat. No. 4,462,046 by Spight discloses an optical vision system in which multiple video cameras in conjunction with a computer are used and multiple off-axis views are stored in a processor. Here, multiple Fast Fourier Transform images (FFT) are video analyzed and computed. This system is limited to a maximum of 30 frames per second.

For Spite's system, the present invention performs optical and parallel processing at speeds close to the speed of light. Many scenes of many different targets can be stored in one compound matched filter, and these filters can be specially designed to handle multiple targets.
A further advantage of the present invention is the specially designed multi
The holographic lens allows for the optical parallel processing of many objects and / or many scenes, and the degree of appearance of the objects can be determined with a desired resolution.

US Pat. No. 3,779,492 to Grumet, which is of interest to the present invention, discloses a matched filter optical correlation system similar to the present invention.
In this system, a coherent optical signal processor is used to recognize a particular known target. Each matched filter recording section has a pair of matched filters for separately processing high and low spatial frequencies. The outputs from them are combined into a logical AND operation, and the targets are examined for exact size and shape and fine features. The matched filter pair optical memory bank has a diffraction pattern of all the analyzable scenes of the target in both azimuth and elevation, i.e. forming the target recognition comb filter bank. ing. All views in this recognition bank are simultaneously optically examined according to the diffraction pattern of the detected object to determine whether the detected object is the desired target stored in any view of the memory bank. Is decided.

"Problems to be Solved by the Invention" The present invention differs from, and evolves from, the Grumte system in several important respects. In one embodiment of the invention, an inverse Fourier transform lens array is used to receive the optical correlation outputs of the matched filter array, each optical correlation output being directed to a separate detector. Gourmet instead shows a single lens 29 that directs all such outputs to a common detector. The invention also provides a normalization means for producing a normalized output signal from the individual matched filters. Gourmet, on the other hand, is completely unaware of the need for signal normalization. The individual detectors described above have the advantage that each output signal can be amplified separately and each amplifier can be used for independent normalization of its processing path (channel). Gourmet also does not recognize that each matched filter has an independent individual angular response curve that can be created and used to determine appearance information about the targeted object.

The general principle of operation of the present invention is that an object is recognized, its size, position and appearance are determined and signals are generated for control. This system recognizes objects,
It takes advantage of the properties of an optical matched filter that allows it to determine its size, its position, its appearance at any angle, and its velocity. It can be done through continuous observation.

In view of these points, the present invention provides a system for optically comparing an input image with optical information stored in one or more matched filters to provide identification and appearance information about the input image. To do.

"Means and Actions for Solving the Problem" An input image is incident on a spatial light modulator, which spatially modulates coherent emitted light. A spatially modulated radiation beam is incident on the multiple holographic lens and a number of Fourier transforms are performed to obtain an array of Fourier transform images of the beam. This array of Fourier-exchanged images is incident on the corresponding matched filter array, each matched filter having a Fourier transform hologram of the outside sight of an object, and its matched
Through the optical correlation signal in the path of the filter, the optical correlation signal depends on the degree of correlation between the Fourier transform image of the spatially modulated radiation beam and the Fourier transform image recorded in the matched filter. Then, the inverse Fourier transform lens means receives a plurality of optical correlation outputs of the matched filter array and performs an inverse Fourier transform on each optical correlation output. Then, the detection means detects the inverse Fourier transform image of each optical correlation output and generates a detection output signal representing each optical correlation output.

According to one preferred embodiment, a plurality of detector output signals are used to determine appearance information about the input image.
It is processed electronically in a processing circuit that compares the relative amplitudes of the signals. The processing circuit includes a normalization amplifier circuit for each detection output signal, an analog / digital converter for converting each normalized detection output signal into a corresponding digital signal, and a comparison for comparing the amplitudes of the plurality of digital signals. It has a circuit. The output of this comparison circuit is processed in a logic circuit that generates a "right movement" or "left movement" robot control signal.

The present invention recognizes the need to normalize the output of individual matched filter paths (channels) in optical correlation systems. The normalization can be done electronically, such as by individual normalization amplifiers to the detectors in each matched filter path. In that case, the gain of each normalization amplifier is set to compensate for the difference in the outputs of the independent matched filters. On the other hand, normalization involves adjusting the amplitude of the light processed through each matched filter path, adjusting the laser output power, and attenuating filters (rotating polarization filters in the optical path of each matched filter path). It can also be performed optically, for example, by providing (usable).

This normalization function is also related to the individual angular response of each matched filter, and the present invention recognizes that each matched filter has its own independent, typically different, angular response curve for each matched filter. It is a thing. Therefore, the angular response curve is empirically determined for each matched filter. The maximum amplitude signals for all angular response curves are set essentially equal to normalize these angular response curves. Then, the angular region of the scene for each normalized angular response curve is determined. This allows for the determination of the number of matched filters required to produce all the angle sensing responses desired for this system.

"Examples" Many elements and concepts relating to the present invention are often used for descriptive purposes, and these are essential for understanding the function and the general operating principle of the present invention. Therefore, the nature and properties of some of these concepts are first described for convenience below.

First, a holographic lens (HL) is created by recording the interference pattern of the emitted light from a diffusing point source and a collimated beam of radiation to create a hologram of the point source. When a holographic lens (after being recorded and processed, such as on a film) is illuminated, it reproduces a point source, ie it behaves like a lens. If the recording process is repeated at the beginning, an array of point source holograms, or what is called a multiple holographic lens (MHL), is recorded on the film.

The present invention provides an array of multiple Fourier transform images of an input spatially modulated laser radiation beam,
Utilizes one of several possible arrangements for offset angle, position and focal length of a multiple holographic lens array. In general, the particular requirements for such arrays will depend on the particular problem of interest. In short, the holographic lens takes out a Fourier transform image (FT) of a scene or a target illuminated by a laser beam, and the multiple holographic lens simultaneously takes out a large number of Fourier transform images. Multiple holographic lens arrays are commonly used in conjunction with corresponding multiple matched filter arrays.

Generally, the lens is spatially modulated (ie,
When illuminated by a collimated beam (such as spatially modulated by passing a scene, object, etc. through the recorded film), the lens produces a Fourier transform image of the object on that film on the focus. This is the basic characteristic of the lens. Then, when the Fourier transform image interferes with a parallel (or reference) beam from the same light source, an interference pattern occurs. This is called a Fourier transform hologram, or matched filter (MF). It is an optical spatial filter of the input object. When an arbitrary scene (scene) is introduced into the system, the matched filter
The created object is extracted and sent out. Then, the signal that has passed through the matched filter is subjected to inverse Lielier transform,
A "correlation" surface (a surface where a signal is detected by a detector or the like) is generated. A sharp correlation signal is generated in the presence of the target recorded in the matched filter, whereas the signal in the absence of the target is a wide and low correlation signal in this correlation plane.

The present invention takes advantage of the matched filter's sensitivity to object rotation or scale size. If any of these appearances change (ie the object is at a different angle or different distance than when the MF was created, ie different scale sizes), the correlation signal changes .

FIG. 1 is a diagram showing a relatively simple configuration example of an optical correlator using a matched filter memory bank according to the technique of the present invention. The target 10 moves past the input to the optical correlator and is imaged (imaged) on the spatial light modulator (SLM) 14 by the input lens 12. The spatial modulator (SLM) 14 spatially modulates the image with respect to the laser beam from the laser 16 which is directed by the mirror 18 and the beam splitter 20. This spatially modulated laser beam is transmitted by the multiple holographic lens 22.
Fourier transform by and directed to the corresponding matched filter array (MFs) 24. Then, the inverse Fourier transform lens array 26 performs an inverse Fourier transform on the output of the MFs 24 and supplies the output to the detector array 28. The output signal is then used to generate an output control signal in element 30.
(The details will be described later).

The matched filter is the size of the input target,
It is a Fourier transform (FT) hologram having characteristics sensitive to the angular direction and the input position. These parameters can be preset to define a set of angle-range (size) boundaries that result in multiple expected object appearances. The detectors can also be provided in separate pieces to analyze the position to the desired degree.

In the production of matched filters, the sharpness of the holographic fringes (hologram stripes) is optimized for the particular spatial frequency that will meet the above size and / or directional sensitivity requirements. There is. Since it is unlikely that both requirements will be met at the same time, multiple independent MFs are used in the present invention. Different specific application properties generally require significantly different MF sensitivities.

The matched filter has a size, wavelength, recording medium thickness, Fourier transform lens focal length, contrast ratio,
It has a complex holographic structure with overlapping, constellation, and spatial frequency components. Each feature can be described using a carefully considered approach for determining and quantifying the quality of matched filters, described below.

(A) A correlation matrix can be considered in the fabrication of a composite target matched filter, and a typical matrix is shown below.

The y-axis is the amplitude transmittance (the amount of transmitted light from 0 to 1.0, which is the inverse of the density). The (relative) exposure is often used as a practical parameter and is shown on the right. The x-axis is the ratio of the reference beam divided by the signal beam and is denoted by the symbol R
Indicated by. In practice, one often determines a beam ratio and then changes the exposure time. This can easily be achieved with a correlator shutter, provided that the procedure is done in the dark. This yields a column value, and a new beam ratio R is set (typically +3 dB or 2) and the procedure is repeated.

The values in the above matrix are obtained by exposing and developing a photographic plate for making a matched filter in an optical correlator,
It is the correlation value of the peak obtained after drying. The development of such a matrix is a time-consuming process and depends on the material, the object, etc. This indicates that making a matched filter with a given material is a complicated and detailed process due to the complex properties of the matched filter.

(B) The standard developed next is called the system standard S.

Where λ (mm) and F (mm) are the operating wavelength of the matched filter and the focal length of the lens, respectively. The smaller the value of S, the easier it is to control the parameter change sensitivity, and the smaller the S, the easier the positional problem. The system criterion S indicates the extent to which the MF spreads on the medium (hence the bandwidth, required exposure time, etc.).

(C) Next, in order to indicate the number of matched filters that can be placed on the photographic plate for recording, a third part called a capacity C is shown.
Standards were set.

Capacitance = Constant × Q · P · G / ((Δν) N · F · λ) ₂ Here, C is the capacity of the matched filter per 1 cm square.

Q is the separated spectrum element (Fourier transform is 180 °
Since it is symmetric, half or all of the matched filter can be used) and takes the value 1 or 2.

P is the number of duplicate filters at a given position. It can take values up to 30, but
Conservatively, it can be said that P is approximately "4".

G is a geometrical packing factor, which depends on the degree of mixing of a plurality of target objects recorded, and
Values between and 1.5 are considered reasonable.

λ and F have already been described. Practical values are 0.45 μm ≦ λ ≦ 2 μm and 25 ≦ F ≦ 1500 cm, respectively.

 N · Δν is the bandwidth of the matched filter.

N is used to obtain 96% of the autocorrelation value Δν
Is a multiplier of. Δν is usually set to a frequency band such that the matched filter is optimized for the holographic contrast ratio.

This description shows the complexity of matched filters, and all of these factors must be taken into account when making matched filters, and when they are complex, the need for accurate normalization. Is meant.

Special equipment was designed to measure the sensitivity of the matched filter to angular orientation, scale and contrast. Furthermore, each of the above criteria needs to be applied to each set of matched filters, which can be, for example,
This is because Δν and angular sensitivity are closely related, and changes in one affect the other. Also, the photo exposure time Δt
Influences Δν and therefore angular sensitivity. Next, consider the value of the system reference S. A change in λ or F, or both, changes S as well as Δν and angular sensitivity. This again illustrates the need for normalization. A similar analysis is possible for changes in scale, or changes in contrast, which requires a reevaluation of the quality of the matched filter.

Another important MF factor is the spatial frequency bandwidth. Matched filters can be optimized for any desired frequency, but the degree of object discrimination is often dependent on the finer parts of the object, the higher frequencies. The frequency requirements need to be considered along with the size, position and appearance of the target. Along the spatial frequency bandwidth, the angle, magnitude and matched
Filter sensitivity also needs to be considered.

FIG. 2 shows a state in which a somewhat rotated Fourier transform image (dotted line) of the same target is directed to an ideal matched filter of the target (solid line). Obviously, there is an overlap between the two in the low spatial frequency bandwidth, which means that some signal for judgment is available. It is clear that if the filter has a high cutoff frequency, it will not have a signal available for robot decision. Moreover, if high frequency filters are used, the central portion of each "curve" section is often omitted.

Next, FIG. 3a shows three angular sensitivity curves obtained for three individual dimensions or spatial frequency bands. When a simulation of the validated correlation process was used, the corresponding correlation curve of Fig. 3b was obtained, indicating the desired result for the target. It can be seen that the angular sensitivity is variable over a wide range.

Another important MF factor is in terms of size. Consider the simple object shown in FIG. The image projected onto the spatial light modulator (SLM) is shown as a system (or front) view. FIG. 4 (A)
The idealized Fourier transform image (FT) for the scene is shown in FIG. The points between the outlined frequency regions are each the zero point of the FT, in this case evenly spaced. When this object brings a new scene, the rectangular image on the SLM has a maximum of 26
Angle (see Fig. 5 (B)), and it is a 90 degree view (see Fig. 5 (C)), that is, its width is (2 · L · COSr + L · SINr: L Lowercase letters, r is the angle shown in the figure), and gradually decreases until the narrowest scene is obtained.

In this process, the FT starts in the state as shown in FIG. As the angular rotation of the scene increases, a larger area produces a similar FT, but with the zero point closer to the origin (zero spatial frequency). Then, at about 26 degrees, each zero point of FT is closest to the origin. In larger angle scenes, the corresponding zeros move to higher frequencies up to 90 degrees, at 90 degrees the largest set of spatial frequencies. This basic example shows that an FT for making an MF needs to be judiciously chosen. When a new scene comes in, the FT "sweeps" over the fixed FT used to generate the FT hologram (matched filter). Then, the correlation signal sequentially takes a series of values until it reaches the maximum value in autocorrelation. Therefore, in the structure of the MF, size factors for multiple targets presenting a common type of scene must be considered. Based on experience, the −3 dB range for the correlation signal can represent a target whose size is distributed from ± 4% to ± 20%.

Then, for a target, -10 °, 0 ° and +
Consider an embodiment similar to FIG. 1 with multiple MFs for three different angle views of 10 °. Their ideal angular sensitivities are shown in FIG. 6 (A). If the scene you encounter is
For example, if it is 0 ° in FIG. 6 (B), the MF corresponding to that scene will generate the maximum autocorrelation signal, while the other two scenes are as shown in FIG. 6 (B). It will only produce a signal that is about half that size. Moreover, the target to be encountered is + 10 ° in FIG. 6C, or − in FIG. 6D.
If 10 °, + 10 ° MF and −10 °, respectively
Maximum signal will be generated from the MF for. In each of these situations, the MF produced at −10 ° and + 10 °
Will not produce a signal, as shown in FIGS. 6 (C) and 6 (D).

This arrangement allows the output signals from the three matched filters to logically determine the angular view of a target. If the target has several other angles within the full view range, three non-obscured sequences of signals are obtained. Multiple paths with more MF can be used for higher accuracy. For a system with 7 channels, the signal would look like the example shown in FIG. The processing logic will be more complicated,
The operating principle is the same. The higher the angular sensitivity of the matched filter, the higher the accuracy, but the larger memory of multiple MFs is required. Generally, the number of MFs required is
It depends on (a) the target and (b) the small part of the target that one wants to analyze. This is probably the case where higher discriminatory power is not needed, since the MF of a target can be made using the low spatial frequency of the target, which results in low directional discriminating ability. On the other hand, high spatial frequencies enhance the system's discrimination power,
This is because the sensitivity of MF rotation is high.

For the exemplary embodiment of the matched filter response of FIG. 6A, the control circuit shown in FIG. 8 can be used for the optical correlation system of FIG. The outputs of the detectors L, C and R are normalized as shown in FIG.
This is because not all MFs have the same perfectly matched autocorrelation signal. The light energy through the MF of one view of the target is also perfectly matched to the second view
It is essentially different from the light energy passing through the MF. For a target with simple symmetry, the three views of the target will generally produce the same autocorrelation. However, for example, three orthogonal scenes of a vehicle require signal normalization to produce quite different autocorrelation signals. FIG. 4 shows an object having different scenes.

In summary, the detected signal is a normalization and gain circuit.
Normalized and amplified at 40. Each of the three views must have the same gain, for example, when the MF for the right view “sees” the signal that the right view provides, but when the MF for the center view “sees” the center view. This is so that the signals will be the same. If the angular response curves are the same, the normalization is that the MF for the central scene has a response of −3 dB for the 0 dB response for the left scene, while
The left-view MF has a −3 dB response to the dB response. ''
Make sure that. The same relationship needs to be true for MF responses in the center and right views. In other words, except for axisymmetric geometric targets, many targets have different cross sections for different (directional) views. Therefore, the energy "passed" through the MF will be different for each different scene, so normalization is required. The normalization factor required for the amplification gain circuit for each matched filter channel is determined from the angular response curve described herein, and the normalization factor for all matched filter channels is stored as data in memory 41. , The data are recalled from the memory 41 to control the amplification gain circuit, respectively.

After normalization, the signals of the R, C and L channels are converted into corresponding digital signals by an A / D (analog / digital) converter 42. The normalized output signal for each channel is
And three main cases facing the left side of the center are shown in FIGS. 6 (B), 6 (C) and 6 (D). Such normalization and A / D conversion can be easily established for intermediate orientations or even for n views of the target.

In FIG. 8, each of the L and R digital outputs is compared with the digital output C in a comparator 44 to obtain a signal C <L,
Generate a set of C> L and C <R, C> R. The following formula can be applied.

(C> R) + (C> L) = center (C> R) + (C <L) = (facing left) facing left (C <R) + (C> L) = (facing right) Rightward, therefore, the appropriate AND gate 46 determines the relationship and generates a move right, stop or move left command.
These commands can be applied to control the robot. However, it should be understood that many other electronic approaches can be used to process the matched filter output described above.

Better angular response can be obtained by the following techniques. That is, the narrower angular sensitivity (ie,
Higher frequency filter), more identification logic,
And a narrower crossover point (eg, a -1 dB crossover portion rather than -3 dB as shown in FIG. 4), or some combination of these three techniques.

The target can also be moved across the image input, which also changes the matched filter output in the correlation plane, resulting in a time-dependent position signal. Therefore, a separate detector could be used to derive the position information.

The embodiments of FIGS. 1 and 8 utilize multiple independent detectors. A television camera can also be used and each line of the sequential scan format containing the correlation signal can be separated and used for the measurement. Although satisfactory, this method has been found to be inaccurate. The correlation surface can also be scanned with a fiber optic probe, which results in a highly accurate measurement of the maximum. However, this technique is manual, slow and not suitable for the present invention. Thus, the output of each independent matched filter path (channel) is directed to a separate detector with balanced output amplifiers so that normalization can be performed according to the nature of each matched filter. There are more advantages.

The exemplary embodiment of the MHL-MF configuration shown in FIG. 9 can provide information about articulation based on the orientation, position and size of the target. If the target types are uniform, the size determination will be used as a range metric. For multiple parameter embodiments, a detector array such as CCD type is required, where the array is subdivided and processed according to those parameters. In the illustrated embodiment, it is divided into nine segments.

FIG. 9 shows a modulated input beam 48, a 3 × 3 multiplex holographic lens 50, a matched filter array 5
2, showing a holographic lens and / or fly-eye shaped MHL 54 and a delimited detector array 56. In this embodiment, the normalization component plate 58 is
It is installed in front of the matched filter plate 52. It consists of a plate large enough to block [(3 × 3) −1] of the beams from the MHL 50 to the MF plate 52. One beam allowed to pass the MF passes through the rotatable polarizer 60. this is,
Linearly polarized sheet-shaped polaroids (HN type) are arranged in a circle. Since the laser beam is linearly polarized, rotation of the circularly arranged polarizer 60 by Θ causes a change of “SIN Θ” in the intensity of the transmitted beam. Therefore, each rotational position can be adapted to an individual non-electronic normalization. The reference beam also
Must be set similarly. In operation, the normalization construction plate 58 will be selectively placed in the x and y directions for each matched filter channel (path). Also, to achieve normalization of all matched filter channels, the rotation of the polarizer will be placed and controlled by the data in memory for each particular matched filter channel. This normalization component plate provides this for each matched filter
For sequential placement with respect to the channels, one can step sequentially in the x and y directions, as with a stepper motor. In addition, the polarizer is MF before the MHL50 or in the series of optical elements in the correlator.
It can be placed at any position that acts on the beam. The technology regarding the installation of the polarizer and its rotation drive is
It is known to rotate the polarizer 60 by driving a stepping motor. The preferred embodiment would be to obtain the control signal for the polarizer from a computer (PC).

Unlike normal neutral density (ND) control that is performed by sensing the laser beam during manufacturing, the polarizer
It should be adjusted by the computer in both the signal and reference beam paths to a predetermined amount derived from the test data of the robot piece through some rotation test.

In some embodiments, the MF correlated robot vision system will require a large MF library. The memory capacity can be determined if the application destination is officially determined and the degree of MF discrimination ability is determined.

The formula of the capacitance C has already been described. With an Argon laser and a 25 mm focal length Fourier transform lens,
1000 M or more per cm ² in the usual conceivable arrangement
It is possible to store F. In particular embodiments and applications, multiple matched filters may be stored for different appearances of different targets.

For many correlator applications, it may be sufficient to obtain a single inverse Fourier transform image of the MF array for processing. Also, in the robot vision system, separate, combed FTs can be utilized to generate multiple independent correlation signals (as shown in FIG. 1). Therefore, it is important that the memory be properly organized. For example, the scale size can be first determined to be a certain size such that a multiple filter bank is formed that simultaneously processes the outputs of multiple filters, and It is also possible to modify the processing system to a suitable configuration such that a single processor may be needed for some scale sizes.

Spatial light modulators (SLMs) are liquid crystal, photorefractive, thermoplastic,
Alternatively, a magneto-optical type can be used and can be either transmissive or reflective. The SLM operates at a few cycles / second, with a contrast ratio of 0.5, a resolution of 50 cycles / mm, non-residuality of the image (no image drift),
It shall have a lifetime of hundreds of thousands of cycles and provide the required 7 or 8 shade levels.

The response of the matched filter can be accurately characterized and can be sufficiently effective through dichromated gelatin or a thermoplastic medium. Also, as an example, multiple MFs with sufficiently high phase frequencies can be computer generated to provide high accuracy.

FIG. 10 shows a configuration for making a target matched filter (MF) 68. In this configuration, the target object 70
Are oriented in a particular direction and are imaged by a lens 72 through a beam splitter 74 onto a spatial light modulator (SLM) 76. Laser light from laser 78 is split by beam splitter 80 into signal path 82 and reference path 84.
The laser beam in the signal path 82 is spatially modulated by the SLM 76, and a Fourier transform image is obtained by the Fourier transform lens 86. The laser beam in reference path 84 is spatially filtered by filter 88 and collimator 90
Are collimated by the beam and directed through the shutter 92 to the matched filter plate 68 where interference occurs and the MF is recorded. When a different orientation of the target is desired, the MF plate is moved to a new position and the above process is repeated. The memory bank of the robot vision system is complete when all appearances of the target have been recorded.
In the practice of this invention, the recording medium may be a photographic emulsion, dichromated gelatin, photopolymer, etc., which can be coated or loaded on a suitable substrate such as a glass plate or a thin film. .

When using the configuration of FIG. 10 to make a matched filter, some precautions need to be taken. First, the lengths of reference path 84 and signal beam path 82 each need to be measured from the center of beam splitter 80. Here, the glass portion in each path and the total length of the glass path, which is increased by the refractive index of the glass (typically 1.5), must be considered. If such considerations are made and the two paths are compared, the total length of both paths must be equal within a difference not greater than the coherence length of the laser, which coherence length is of the order of a few centimeters. In practice, a difference of 2 millimeters can easily be realized.

Based on the focal length of the Fourier transform lens 86 (360 mm in a typical embodiment) and the operating wavelength (typically 6328 Å), the system factor S given in advance can be calculated to be 4.39 (cycles / mm) / mm. Becomes
A cutoff spatial frequency of 10 cycles / mm can be used with a predetermined representative standard. Therefore, the center-to-center distance for the matched filter spacing in the array illustrated in FIG. 9 is (2 ×
10) /4.39=4.56mm.

A second solution to the configuration of FIG. 10 is to use openings during the fabrication of the array that block the fabrication of one filter while blocking all other filters. Then, this aperture is moved to the next place by 4.56 mm, and the second filter is made. At the same time, this movable opening has a diameter of 4.56 mm. Figures 3 and 4 of U.S. Pat. No. 4,703,944 show a device that allows to hide all but one target position.

Third, the elements of FIG. 10 need to be arranged with particular care so that the image of the target of interest is focused on the spatial light modulator 76 at the input of FIG. To read the spatial light modulator 76, a polarizer 94 placed in front of the SLM, and
An analyzer 96 placed behind is required. These three elements need to be arranged individually and in tandem so that the output from them has the same polarization as the reference path beam. To make a holographic matched filter, the reference and signal beams must be in the same plane of polarization for maximum effect. Any other arrangement will degrade the quality of the matched filter, and even orthogonal polarization states will not even result in holographic interactions.

The Fourier transform lens 86 may be a glass lens, but a specially designed multiple holographic lens is preferably used to direct the numerous FT replicas of the target to all desired MF positions. When the holographic lens is illuminated with a collimated beam of radiation, an off-axis focus is obtained. If the beam is varied in wavelength while remaining parallel, a second off-axis focus with a different offset angle and focal length than in the first case is obtained. This result is physically due to the fact that holograms are basically highly complex diffraction gratings. It is often beneficial to fabricate a matched filter at one wavelength and use that filter at another second wavelength. For example, some images are best recorded on matched filters at wavelengths in the blue light spectral band and best reproduced at wavelengths in the red spectral band. In this case, the acting optical signal tends to change the image formed on the matched filter. This trend is
If the filter is used at the wavelength used during fabrication, it will be significantly reduced.

In addition, the outside sights with different targets are uniquely matched.
It should be appreciated that the filter results and the filter itself has angular scale sensitivity. In a robot vision system, each different appearance matched filter should be treated like a matched filter for some new object, because each matched filter is different and produces a different signal. is there. Therefore, the same standard is required, i.e. normalization for a properly aligned matched filter. Further, the normalization is affected by the angular sensitivity and thus requires correction, which is not recognized in the prior art.

For example, for a robot system for "LCR" movement, matched filters must be made for each of the left (L), center (C), and right (R) views. And to produce the actual used normalized response curve for each matched filter,
An angular sensitivity curve has to be generated. The angular response curve cannot be made for the central position of the scene and used for the left and right scenes, which is
This is because the angular response curves for each matched filter are different.

The number of MFs required depends on the response of the individual filters. For example, a particular target for which a matched filter is made may exhibit an angular response as shown by one of the rounded protrusions in FIG. If this is the view from above and the target always looks the same when rotated, then each response will be the same, and the memory for the 360 ° full range will be as shown in FIG. fence)"
Will have a type response. Note that the responses are all equally spaced and of equal height.

Here, consider a moving robot (machine) unit. If this robot part is considered from a plurality of different outdoor sightseeing scenes, the set of responses shown in FIG. 12 will be obtained, for example.
Each response may be a 3dB response (ie the baseline lies at half the peak level), but it should be noted that the magnitude of the angular width will change and the peak correlation signals will have different heights. is there. These are different outlooks
Due to the fact that one scene may have more energy than another.

The end result is that their amplitudes need to be brought to some common level. This can be done electronically by a normalizing amplifier as described above, or by multiple matched
Filtered photographic plates can be accomplished optically by selectively exposing them to "blacken" uniformly, reducing their amplitude to some common maximum height.

And the resulting response would be as shown in FIG. Numbers represent arbitrary values and are for illustration only. However, it should be noted that the peak values are equal. The number of matched filters required for a particular angular range can be determined after the angular response has been normalized as shown in FIG.

While the embodiments of the present invention and modifications thereof for a robot vision system have been described in detail above, it should be apparent that the disclosure and technology of the present invention suggest many modifications to those skilled in the art. is there.

[Brief description of drawings]

FIG. 1 is a diagram showing the configuration of an exemplary embodiment of a robot vision system for teaching and explaining the overall principle of the present invention, and FIG. 2 illustrates the angular sensitivity of a matched filter with respect to spatial frequency width. Figures 3a and 3b illustrate the sensitivity of the matched filter to image rotation at a given scale size, and Figure 4 shows a simple object with rotation of appearance in plan view. 5 and 5 are views for explaining the influence of the rotation of the object shown in FIG. 4 on its Fourier transform image, and FIG. 6 is a view showing the angular sensitivity of the matched filter. FIG. 7 is a diagram showing the angular sensitivities of correlation signals of three matched filters with respect to various object positions. FIG. 7 is a typical example of the correlation signals in a system in which seven kinds of angular scenes are recorded on a plurality of matched filters. FIG. 8 is a block diagram showing the configuration of a position control system having signal processing electronics for obtaining position control signals, and FIG. 9 is another implementation of the invention having a plurality of divided detector arrays. FIG.
FIG. 10 is a diagram showing one optical technique for normalizing the output of each matched filter path, and FIG. 10 is a diagram showing a typical configuration example for making an MF memory bank for a plurality of different outdoor sightseeing scenes with respect to a target object. , Figure 11 shows all matched
FIG. 12 shows an ideal situation where the angular response curves of the filters are equal in their amplitude and angular range. FIG. 12 shows a practical case where the angular response curves of each matched filter differ in both their amplitude and angular range. Figure 13, Figure 13, shows the normalization of the response curve of Figure 12, showing that the maximum amplitude of each curve is equal, resulting in different angular ranges for each matched filter. is there. 14 ... Spatial light modulator, 22 ... Multiple holographic lens, 24 ... Matched filter array, 26 ... Inverse Fourier transform lens array, 28 ... Detector array, 40 ... Normalization / gain circuit, 41 …… Memory, 42 …… A / D converter, 4
4 …… Comparator, 46 …… AND gate

Claims (10)

[Claims] 1. An optical correlation system that optically compares an input image with optical information recorded on a matched filter to provide identification and appearance information of the input image, wherein: A spatial light modulator that is incident and spatially modulates a parallel radiation beam by the image to form a spatially modulated radiation beam; (b) the spatially modulated radiation beam is incident, A multiple holographic lens that performs a number of Fourier transforms to obtain an array of Fourier-transformed images of the radiation beam, and (c) a matched filter array upon which the array of Fourier-transformed images is incident, each matched filter array The filter has a Fourier transform hologram of the out-of-view scene of a target and passes the optical correlation signal in the path of its matched filter, the optical correlation signal A matched filter array that is dependent on the degree of correlation between the Fourier transform image of the modulated radiation beam and the Fourier transform image recorded on the matched filter; and (d) a plurality of matched filter arrays of the matched filter array. An inverse Fourier transform lens means for receiving the optical correlation output and performing an inverse Fourier transform on each optical correlation output, and (e) detecting an inverse Fourier transform image of each optical correlation output, and outputting a detection output signal representing each optical correlation output. Detecting means for generating, and (f) normalizing means for generating a normalized detection output signal for the path of each of the matched filters, the normalization creating an angular response curve for the individual matched filters. Then, normalize these angular response curves so that the maximum amplitude signals of all angular response curves are essentially equal, and each normalized angle Normalizing means including determining an angle region of the scene with respect to the answer curve, and (g) comparing the magnitudes of the plurality of normalized detected output signals, and the appearance information of the input image. And a comparing means for generating a plurality of output direction control signals defined by the optical correlation system. 2. An optical correlation system according to claim 1, wherein the input image is optically compared with the optical information recorded in the matched filter to provide identification and appearance information of the input image. The array of matched filters includes at least a central matched filter for a view on a central axis of the target, a left matched filter for a view to the left of the central angular view of the target, and a central angular view of the target. A right matched filter for the right scene, (b) said detection means for at least a central detector for said central matched filter, a left detector for said left matched filter, and for said right matched filter A right detector, and (c) the comparison means is at least the magnitude of the output signal of the left detector and the central detector. The left comparing means for comparing these to determine which of the magnitudes of the output signals is greater, and which of the magnitude of the output signal of the right detector and the magnitude of the output signal of the central detector are greater, And a right comparing means for comparing these to determine
(D) directing means for receiving respective outputs of the left comparing means and the right comparing means and for generating an output direction control signal representing the determined angular position of the target; and (e) the detecting means. Processing means coupled to receive the plurality of detection output signals and comparing the relative magnitudes of these signals to determine appearance information of the input image, the processing means further comprising: Normalization circuit for detection output signal, analog / digital converter for converting each normalized detection output signal to corresponding digital signal, and comparison circuit means for comparing amplitudes of the corresponding plurality of digital signals The optical correlation system according to claim 1, further comprising: 3. An optical correlation system according to claim 1, wherein the input image is optically compared with the optical information recorded in the matched filter to provide identification and appearance information of the input image. The optical correlation system according to claim 1, further comprising a normalized amplifier circuit that generates a normalized detection output signal for each of the detection output signals. 4. The optical correlation system according to claim 1, wherein the input image is optically compared with the optical information recorded in the matched filter to provide identification and appearance information of the input image. The optical correlation system of claim 1, wherein the optical correlation system comprises an optical normalizer in the path of each matched filter. 5. An optical correlation method for optically comparing an input image with optical information recorded on a matched filter to provide identification and appearance information of the input image, the method comprising: Leading to a spatial light modulator, spatially modulating a parallel radiation beam to form a spatially modulated radiation beam, (b) allowing the spatially modulated radiation beam to enter a multiple holographic lens A plurality of Fourier transforms to obtain an array of Fourier transform images of the spatially modulated radiation beam, and (c) injecting the array of Fourier transform images into a matched filter array. Then, each matched filter has a Fourier transform hologram of an outside scene with an object and passes the optical correlation signal in the path of the matched filter. The optical correlation signal is dependent on the degree of correlation between the Fourier transform image of the spatially modulated radiation beam and the Fourier transform image recorded in the matched filter, and (d) the matched filter array A plurality of optical correlation outputs are made incident on the inverse Fourier transform lens means, an inverse Fourier transform is performed on each optical correlation output, and (e) an inverse Fourier transform image of each optical correlation output is detected to detect each optical correlation output. Producing an output signal, and (f) normalizing the signal passing through the path of each said matched filter to produce a normalized detection output signal, the normalization step relating to the individual matched filters. , Creating an angular response curve having independent maximum amplitude and angular range, and normalizing the signal passing through the path of each matched filter. And equalize the maximum amplitude of all the angle response curves,
This results in the generation of different angular ranges for each matched filter path, and after making the maximum amplitude signals of all angular response curves essentially equal to normalize the angular response curves, each normal Determining the angular region of the scene with respect to the normalized angular response curve, and (g) comparing a plurality of the normalized detected output signal magnitudes with the appearance information about the input image. An optical correlation method comprising: each step of generating a plurality of output direction control signals that are defined. 6. An optical correlation method according to claim 5, wherein the input image is optically compared with the optical information recorded in the matched filter to provide identification and appearance information of the input image. The step of making the array of Fourier transform images incident on the matched filter array comprises the steps of setting the array of Fourier transform images to at least a central matched filter for a view on the central axis of the target and a central angular view of the target. Including the step of causing the left matched filter for the left scene and the right matched filter for the scene to the right of the central angle scene of the target to be incident, and (b) at least the inverse Fourier transform image of the central matched scene. A center detector for the filter, a left detector for the left matched filter, and a right matched filter for the left matched filter.
(C) The magnitude of the normalized plurality of detection output signals is at least the magnitude of the output signal of the left detector and the magnitude of the output signal of the central detector. Left comparing means for comparing them to determine which is greater, and these to determine which of the magnitude of the output signal of the right detector and the magnitude of the output signal of the central detector is greater. And (d) each output of the left comparing means and the right comparing means is used to generate an output direction control signal representing the determined angular position of the target, ) The process of determining the appearance information of the input image by comparing the relative magnitudes of the plurality of detection output signals, in which a normalization circuit for each detection output signal is associated with each normalized detection output signal. Convert to digital signal An analog / digital converter of order, the corresponding optical correlation method according to claim 5, characterized by having a process and a comparison circuit means for comparing the amplitudes of a plurality of digital signals. 7. An optical correlation method according to claim 5, wherein the input image is optically compared with the optical information recorded in the matched filter to provide identification and appearance information of the input image. The steps are performed by a normalization amplifier circuit that produces a normalized detection output signal for each of the detection output signals, to obtain a normalized detection output signal for all of the matched filter paths. 6. The optical correlation method according to claim 5, further comprising the step of storing an amplification factor for each of the normalized amplification circuits in a memory. 8. An optical correlation method according to claim 5, wherein the input image is optically compared with optical information recorded on a matched filter to provide identification and appearance information of the input image. 6. The optical correlation method according to claim 5, wherein the step is performed by optical normalization in the path of each matched filter. 9. An optical correlation method according to claim 8, wherein the input image is optically compared with the optical information recorded on the matched filter to provide identification and appearance information of the input image. The process is performed by an optical attenuating filter in the path of each matched filter, and the normalization process is performed by a rotatable polarization attenuating filter in the path of each matched filter. And wherein the rotatable polarization-attenuating filter is selectively rotated for each matched filter path to perform normalization on all of the matched filter paths. Optical correlation method. 10. An optical correlation method according to claim 8, wherein the input image is optically compared with the optical information recorded in the matched filter to provide identification and appearance information of the input image. 9. The optical correlation method according to claim 8, wherein the step is performed by controlling the power of laser light passing through each of the matched filter paths.