Body Scanner with Reference Database
1. A method for performing automated target recognition on a body scanner, comprising:
- acquiring database scans;
converting said database scans to humanoid coordinates;
acquiring subject scan;
converting said subject scan to humanoid coordinates;
comparing said subject scan to said database scans;
identifying features in said subject scan that appear infrequently in said database scans.
Body scanners are used in airports and other secured facilities to detect weapons, explosives, and other security threats hidden under persons'"'"' clothing. These devices use x-rays, millimeter waves and other radiant energy to produce an electronic image of the person'"'"'s body and any concealed objects. Examination of these images by human analysts is slow, expensive, and subject to privacy concerns. The Invention provides automated analysis by comparing each image against a database of previous scans, using a plurality of subjects with different body types. This comparison is facilitated by digitally mapping each body scanner image to humanoid coordinates. This overcomes the failings of the prior art by allowing each anatomic location on one person to be referenced to the same anatomic location on all other persons.
- 1. A method for performing automated target recognition on a body scanner, comprising:
acquiring database scans; converting said database scans to humanoid coordinates; acquiring subject scan; converting said subject scan to humanoid coordinates; comparing said subject scan to said database scans; identifying features in said subject scan that appear infrequently in said database scans.
- View Dependent Claims (2, 3, 4, 5, 6)
- 7. An apparatus for searching a person entering a security controlled area, comprising:
a body scanner producing an electronic image of the person'"'"'s body; a digital computer, said digital computer comparing said electronic image against a database to detect uncommon image features in said electronic image; said comparing comprising a spatial frame-of-reference matched to the human body; and an annunciator responsive to the detection of said uncommon image features.
- View Dependent Claims (8, 9, 10, 11, 12)
- 13. A method for detecting security threats on a person, comprising:
first operation of a first body scanner to produce a database;
said database comprising common features of the human body;
second operation of a second body scanner to produce a subject image;
said subject image comprising features of the person'"'"'s body;
comparing said subject image against said database in humanoid coordinates.
- View Dependent Claims (14, 15, 16, 17, 18)
This application claims the benefit of provisional patent application No. 62/406,702 filed with the USPTO on Oct. 11, 2016, with the same title and inventor.
This invention relates to screening persons for security threats through the automated analysis of body scanner images.
A variety of body scanners have been developed to detect weapons, explosives and contraband concealed under the clothing of persons entering security controlled areas. These devices operate by detecting radiant energy that has been modulated by or emitted from the body of the person being examined. Radiant energies used include: x-rays, microwaves, millimeter waves, infrared light, terahertz waves, and ultrasound. In a typical operation, the person being examined is actively exposed to a beam of millimeter waves or x-rays. A portion of this radiant energy interacts with the person, their clothing, and any concealed objects they may be carrying. This interaction modulates the radiant energy reflected from, scattered by, or transmitted through the person. This reflected, scattered or transmitted radiant energy is collected by sensitive detectors, and the resulting electronic signals are routed to a digital computer. Alternatively, some body scanners operate passively, collecting radiant energy that has been thermally emitted from the person'"'"'s body and surrounding area. Examples of this are infrared and millimeter wave sensitive cameras. Regardless of active or passive operation, body scanners convert the electronic signals from their detectors into digitally represented images of the person'"'"'s body. In these images the clothing is essentially transparent, allowing the security officer to visualize objects that are concealed underneath the clothing. Commercial body scanners include the model AIT84, sold by Tek84 Engineering Group, San Diego, Calif.; model SECURE 1000, sold by Rapiscan Security Products, Torrance, Calif.; model SmartCheck, sold by American Science and Engineering, Billerica, Mass.; model ProVision, sold by L-3 Communications, Woburn, Mass.; and model Eqo, sold by Smiths Detection, Edgewood, Md.
In spite of using different radiant energies and imaging geometries, body scanners detect concealed objects in the same fundamental way: they create an electronic image of the person with the clothing being essentially transparent. This electronic image is composed of bits of digital information, which may reside in a storage medium, computer processing unit, or other device capable of retaining such data. For image storage this may be done in a common file format, such as jpg, bmp, or gif. Within a computer processing unit the storage will typically be pixel values ordered in a row and column arrangement. The electronic image can be manipulated directly in digital form, or converted to a visual image by devices such as image printers and video monitors. As used here, and commonly in the art, the term “image” and “scan” refer to the bits of digital information residing in a digital device, the visual display of this information on a video monitor, the printed image corresponding to this digital information, and other such data presentations. These concepts of digitally representing and manipulating images are well known in the art of image processing.
All body scanners incorporate a digital computer, as shown in
Body scanners are capable of detecting a wide range of security threats and contraband; however, the required image interpretation by the Security Officer presents a multitude of problems and difficulties. The manpower requirement to operate these systems is very high, requiring a security officer to be present to analyze each image. This is aggravated by the specialized training each security officer must go through to be proficient in the image analysis. The inherent limitations of human image analysis, and the nature of this work, promotes errors and deficiencies in the security screening process. For instance, security officers may become distracted or tired and miss concealed objects. In a worse scenario, a security officer may be bribed or coerced to ignore concealed objects. Further, human image analysis requires that the examination area be large enough to accommodate the extra security officer. Further, humans require about ten seconds to process each image, which can slow the throughput of the security checkpoint. Still further, some persons object to an electronic image of their unclothed body being displayed and viewed by the security officer.
The performance of ATR can be statistically measured in terms of the probability of detecting certain types of concealed objects, versus the false alarm rate. A well performing system detects a high percentage of the concealed objects with minimal false alarms. Conversely, a poorly performing system has a low detection probability, and a high false alarm rate. Humans perform exceedingly well at this task with a seemingly trivial effort. This can be appreciated simply by looking at the high-quality body scan image 70 in
Prior Art approaches generally employ one of two strategies for ATR. The first uses reference information that is contained within the image being examined. One such scenario is described in U.S. Pat. No. 8,194,822, to Rothschild et al., issued Sep. 28, 2010. In this '"'"'822 Invention, the value of a particular pixel is compared against a plurality of other pixels values in the same image, providing a reference to detect uncommonly bright or dark areas. Another Invention using this first general strategy is described in U.S. Patent Application No. 62/406,702, to Smith, filed Oct. 11, 2016, and incorporated herein by reference. In this Invention, an image feature at one location in the image is compared to the bilateral location in the same image, thereby detecting features that are bilaterally symmetric. In general, human anatomy is bilaterally symmetric, while concealed threats are bilaterally asymmetric. Therefore, this provides a method for separating anatomic from threat features, by only using information contained within the single scanned image.
The second general approach to ATR is to use reference information in a library, i.e., a database, of previously acquired scans. This strategy is described in U.S. Pat. No. 5,181,234, to Smith, issued May 22, 1991, and incorporated herein by reference. In the '"'"'234 Invention, a plurality of body scanner images is acquired and analyzed for the presence of image features. These features are stored in a library for later comparison to scans of actual subjects. Features appearing in the actual scans, which do not appear in the library, are classified as potential threats. As briefly stated in the '"'"'234 patent: “The location of detected features can be referenced to the absolute location in the image, or in relation to the body of the person being examined” (col. 14 lines 41-43). However, '"'"'234 is silent on how a location “in relation to the body” can be calculated or otherwise determined.
Inadequate ATR has placed severe limitations on the use of body scanners. Security personnel at airports, military bases and Government facilities have been faced with undesirable alternatives. One alternative is to use body scanners with human image analysts, providing excellent detection capability and few false alarms. However, they also must accept the associated manpower problems, long analysis times, and privacy concerns. The other alternative has been to use body scanners with prior art ATR. This provides high-throughput, reduced personnel requirements and far better privacy to the person being screened. However, in this alternative, the primary purpose of the body scanner is largely defeated, a result of the poor detection probability and frequency false alarms. A third alternative, which is often selected, is to not use body scanners because of the unacceptable problems of either using, and not using, prior art ATR. Indeed, the performance of ATR is the critical factor in the widespread use of body scanners in security facilities.
The Present Invention provides body scanner automation by comparing the scanned image against a database of previously scanned images in a common spatial frame-of-reference matched to the human body, referred to as humanoid coordinates. These humanoid coordinates enable each feature in the scanned image to be classified according to how common it is in the general population of persons being scanned. Features that are common are ignored. Features that are uncommon are suggestive of a threat object being present, and subjected to further scrutiny.
In one preferred embodiment, a body scanner is used to acquire a database of scanned images, representative of the general population of persons that may be searched. A computer algorithm converts the images in this database from the spatial coordinates produced by the scanner to humanoid coordinates. The goal and purpose of the humanoid coordinates is to align the body parts appearing in each of the images. That is, the arms, head, torso, legs and feet in any one image are spatially aligned with the arms, head, torso, legs and feet in all of the images. Subsequently, the body scanner is used to acquire a scanned image of a subject that is being screened for the presence of concealed threats. This subject image is also converted to humanoid coordinates. Image features are identified in the subject scan, such as the brightness of an area, the strength of an image edge, the angle of an image edge, a measure of the image texture, detection of geometric shapes such as corners, or other characteristics known to those in the field of digital image processing. Each of these features appearing in the subject scan is evaluated for its frequency of occurrence in the database scans. This is accomplished by searching each image in the database for the presence or absence of the indicated feature, at the same location in humanoid coordinates that the feature appears in the subject scan. Image features in the subject scan that appear commonly in the database scans are classified as normal and discarded from further consideration. Conversely, image features in the subject scan that do not appear in the database scans, or only infrequently occur, are classified as abnormal or suspicious, and further evaluated. In another preferred embodiment, the database scans are consolidated into a statistical representation to facilitate their comparison to the subject scan.
It is therefore a goal of the Present Invention to provide an improved method and apparatus for detecting security threats concealed under the clothing of persons entering a security controlled area. Another goal of the Present Invention is to provide improved ATR capability for body scanners. Yet another goal is to use the information embedded in a database of scans to represent the characteristics of the general population that may be searched with the body scanner. Still another goal is to facilitate the comparison of body scanner images through a humanoid coordinate system. A further goal is to provide ATR discrimination based on the probability of occurrence of features in the scanned image. Yet another goal is to eliminate the need for humans to view and analyze body scanner images.
In a second step, the primary fiducial markers 225-246 are identified on the outline 221. These are located through common image processing algorithms looking for specific image features. In a preferred embodiment, the wrists 225 226 227 228 and ankles 241 242 243 244 are defined by locating the narrowest point across the limb. The inside and outside of the elbows 229 230 231 232, and the tips of the feet 245 246 are identifiable by the abrupt change in slope of the outline 221. The neck 234 235 and groin 239 are readily located as the lowest and highest points in the local region, respectively. The armpits 236 237 are determined by starting at the neck fiducials 234 235, respectively, and moving outward and down until intersecting the outline 221. Likewise, the hip fiducials 238 240 are at the location on the outline 221 with the same height as the groin 239. The top of the head 233 is located by finding the best-fit circle matching the top of the head, then constructing a line between the center of this circle and the midpoint between the neck fiducials 234 235. The top of the head 233 is then identified as the point where this line intersects the outline 221. Algorithms to carry out these steps are routinely known in digital image processing, have many variations, and are tailored to the particular type of body scanner being used.
The third step shown in
A key feature of this multitude of fiducials is that they occur in bilateral pairs. For example, the two armpit fiducials 236 237 form such a pair. As shown in the data representation 280, the midpoint 282 between these fiducials 236 237 is located on the body'"'"'s axis of symmetry 203 of the body outline 221. Likewise fiducials 255 and 258 form a pair around midpoint 283, and fiducials 270 271 form a pair around midpoint 284. Put in other words, the vertical axis of symmetry 203 of the body outline 221 can be calculated as all of the midpoint locations [e.g., 282 283 284] of all the pairs [e.g., 236 and 237, 255 and 258, 270 and 271, respectively].
The Present Invention is based on conversion from the scanned image coordinates to humanoid coordinates.
This new spatial coordinate system, referred to as humanoid coordinates, is based on the characteristics of the human body, thereby facilitating the comparison of images of the human body. It is a fixed, single, common coordinate system for all scanner images involved in the Present Invention. In the preferred embodiment of the Present Invention it is formatted to resemble a simplified outline of the human body. That is, every image generated from the body scanner will be unique in its grayscale representation, body outline, and fiducials, as typified in
The purpose and goal of this coordinate conversion is to spatially align all scanner images with each other, allowing a comparison among them. Without this coordinate conversion it is difficult or impossible to compare any two images, much less a multitude of images, because of the infinite variation in body shapes and postures. Accordingly, the specific characteristics of the humanoid coordinates are not important; only that the process is the same for all images, and the humanoid coordinates generally resemble the human body. It is within the scope of the Present Invention that the humanoid coordinates could be as diverse as tabulated data in a spreadsheet, to rectangular body outlines in an image format, to formats derived from known techniques such as principle component analysis.
In further detail, the body outline with fiducials 290 is calculated by finding the outline of the body, locating key landmarks on the body as primary fiducials, inserting a selected number of secondary fiducials between the primary fiducials, and connecting the corresponding fiducials to form quadrilaterals. In the humanoid coordinates this procedure also defines a body outline with fiducials 390, with a one-to-one correspondence to the original image 290. That is, the exemplary fiducials in the scanned coordinates 270-273 correspond to fiducials 370-373, respectively, in the humanoid coordinate frame. Likewise, fiducials 229-232 correspond with fiducials 329-332 and fiducials 255 283 258 256 285 259 245 246 correspond to 355 383 358 356 385 359 345 346, respectively. The exemplary quadrilateral defined by fiducials 272 273 232 231 thereby corresponding to the quadrilateral defined by fiducials 372 373 332 331.
The bilinear transform, a mathematical mapping technique known in the field of digital image processing, maps the location of each point within one quadrilateral into a corresponding location in a second quadrilateral. In a preferred embodiment, a computer routine loops through each of the pixel locations in the humanoid coordinate image 396. In
As thus shown, any image 204 produced by the body scanner can be converted into an image 396 in humanoid coordinates.
In one embodiment, the feature under consideration in the scanned image is sequentially compared against the corresponding locations in each image in the reference database. This procedure determines the fraction of images in the database that reasonably contain the feature. For example, the feature of the shin 420 is reasonably matched in two of the database regions 520 522, and not matched in the remaining two regions 521 523. Therefore, this feature is contained in 50% of the database images. In the second example, the metal object on the belt feature 410 is contained in none of the corresponding regions in the database 510 511 512 513. Simply thresholding this frequency of occurrence parameter therefore results in the shin 420 being classified as not a threat, while the metal object on the belt 410 is classified as a threat.
In a preferred embodiment, the database scans are consolidated into a statistical representation to facilitate their comparison to the subject scan. As an example, the image feature being evaluated may be the grayscale value of each individual pixel in the subject image. In images created by backscatter x-ray body scanners, for example, dark pixels may indicate the presence of metal objects, while bright pixels may indicate the presence of explosives. However, ATR cannot be achieved by simply thresholding the subject scan for dark or bright pixels, since the normal variations in the human body also result in pixels that are dark and bright. In this embodiment, all of the images in the database are analyzed to find the mean pixel brightness, and the standard deviation of the pixel brightness, at each location in the humanoid coordinates. Further, this may be formatted as two grayscale images, where each pixel value in the first image is the mean, and each pixel value in the second image is the standard deviation. In this manner the multitude of images in the database are consolidated into only two images, which contain the key information for comparing the parameter of image brightness.
This consolidation of the database into mean and standard deviation images has the further advantage of allowing the ATR algorithms to be described as array processing, rather than pixel-by-pixel operations. In a preferred embodiment, a “standard deviation above the mean” (SDAM) image is created by subtracting the mean image from the subject image, and dividing the result by the standard deviation image. As known in the art of array processing, this defines that all pixels in the images are processed in the indicated manner. Each pixel value in the SDAM image is therefore a statistical measure of the typicality of the grayscale value at this location in the subject scan image, taken in comparison to the database scans. That is, for example, a value of −3 at a particular location in the SDAM image would indicate that the corresponding location in the subject image is exceedingly dark, three standard deviations below the mean in brightness, compared to the database images. As known in statistics, a numerical value more than three standard deviations from the mean occurs in only about 1 in 300 random occurrences. Therefore, this pixel value in the subject image would be highly uncommon, and indicative of a metal object. Likewise, a value of +3 would be exceedingly bright, three standard deviations above the mean, and indicative of explosive material. Pixel values in SDAM between about −2 and 2 are exceedingly common, and can immediately be dismissed from consideration. Again using array processing terminology, ATR detection of metal threats is accomplished by thresholding the SDAM image for values below a specified threshold, typically −2 to −5. Likewise, ATR detection of explosives is accomplished by thresholding the SDAM image for values above a specified threshold, typically 2 to 5.
In an analogous manner, a preferred embodiment of the Present Invention calculates and thresholds SDAM images for a variety of other features besides pixel brightness. This is essentially equivalent to consolidating the reference database in different fashions to facilitate specific comparisons. For instance, this includes edge sharpness, which can be calculated at each point in the subject image as the magnitude of the image gradient. Extended further, separate SDAM images of edge sharpness may be generated and evaluated for different edge angles. Further, the consolidation of the reference database may provide separation according to the sex of the subject, if the scan is a front or rear view of the subject, the subject'"'"'s body type, and so on. Those skilled in the art will recognize that these are specific embodiments of comparing features in the subject scan with the database scans to identify those features that are uncommon.
As thus described, the images from the body scanner are converted into humanoid coordinates to facilitate the comparison of image features between images. This can be carried out by converting the entire scanned image into humanoid coordinates, and subsequently performing the feature detection and comparison operations on the converted data. In another preferred embodiment the feature detection is performed before the conversion operation. That is, features such as edges and brightness variations are identified in the image produced by the body scanner. The location of these features is then converted into humanoid coordinates, thereby enabling the comparison of these features between images. That is, the order of operations can be reversed, while still achieving the same result.
Although particular embodiments of the Present Invention have been described in detail for the purpose of illustration, various other modifications may be made without departing from the spirit and scope of the Invention. Different methods for identifying the outline of the body and fiducial location may be used. Likewise, different digital image warping techniques may be used to convert the images to humanoid coordinates. The format of the humanoid coordinates may be in other image and non-image formats. The reference database may be stored in other formats or statistical consolidations. The data representations at the various steps in the embodiments may be discrete, such as pixel values in a digital image, or mathematical, such as equations representing curves, or mathematical interpolations between discrete values. The computational platform to carry out the algorithms of the Present Invention may be a conventional sequential instruction computer, or a parallel hardware device such as an FPGA. The Present Invention may be used in conjunction with complementary ATR algorithms such as asymmetry detection and neural networks.