EP3103060A1 - 2d image analyzer - Google Patents

Abstract

The invention relates to a 2D image analyzer comprising an image scaler, an image generator, and a pattern finder. The image scaler is designed to scale an image according to a scaling factor. The image generator is designed to generate an overview image which has a plurality of copies of the received and scaled image. Each copy is scaled by a different scaling factor. In the process, the respective position can be calculated using an algorithm which takes into consideration a distance between the scaled images in the overview image, a distance from the scaled images to one or more of the boundaries of the overview image, and/or other predefined conditions. The pattern finder is designed to carry out a characteristic transformation and classification of the overview image and to output a position where the searched pattern maximally matches the specified pattern. Optionally, a post-processing device can also be provided for smoothing and correcting the position of local maxima in the classified overview image.

Description

 2D BildanaIysator

description

Exemplary embodiments of the present invention relate to a 2D image analyzer and to a corresponding method.

In many cases, digital image processing uses multiple scaled versions of the input image for pattern recognition. An example of such a pattern recognizer is the classifier by Viola-Jones, which is trained on a specific model size.

Depending on the distance that the desired pattern (eg face or eye) has to the camera system, a larger or smaller image is displayed. The fixed model size of the classifier therefore has to be recorded in several scaling levels

Image in order to obtain an optimal match with the model size of the classifier in one of the scaling levels. The scaling levels are usually searched in ascending or descending order (see picture pyramid). This sequnetial execution is very poorly suited, in particular, for parallel architectures (eg FPGA).

Therefore, there is a need for an improved approach. The task is to enable an efficient and reliable recognition of a pattern.

The object is solved by the independent claims. Further developments are defined in the subclaims. Embodiments provide a 2D image analyzer having an image calibrator, an image generator, and a pattern finder. The image calibrator is configured to receive an image having a searched pattern and to scale the received image according to a scaling factor. The image generator is configured to generate a survey image having a plurality of copies of the received and scaled image, each copy scaled by a different scaling factor. The pattern finder is configured to compare a predetermined pattern with the plurality of received or scaled images within the overview image, and to output information regarding a position in which a match between maximum is the searched pattern and the predetermined pattern, the position relating to a respective copy of the received and scaled image.

The core of the present invention is therefore that the recognizability of certain patterns changes over the image size. In particular, it has been recognized here that the sequential execution of subschield-sensitive scaling stages can be very inefficient. Therefore, embodiments of the present invention provide a bi-directional analyzer with a pattern finder, wherein the pattern finder is applied to an overview image that includes the image in which the searched pattern is contained at different scaling levels. By means of these different scaling stages-which are skilfully arranged in an over-picture picture-it can be achieved that in a single step, several or all scaling stages of interest can be searched at once (and not, as usual, pattern recognition is applied to several scaling stages in succession). , This happens because the pattern recognition only takes place on the over-image. Where in the overview image the largest match with the searched pattern is present, it is clear, on the one hand, which scaling level this match has delivered (position of the scaled image within the overview image) and, on the other hand, at which position (x, y coordinate in the image) this was (Position within the scaled image in the overview image, corrected by the scaling factor). The bottom line is that it offers the advantage of being much more efficient, especially on parallel architectures such as e.g. FPGAs, the searched pattern can be detected in at least one of the scaling levels by only pattern recognition on the overview image. By knowing the scaling, the position of the desired pattern can then also be calculated in the absolute image.

Therefore, according to further embodiments, each scaled image according to the respective scaling factor is associated with a respective position in the overview image. In this case, the respective position can be calculated by an algorithm which takes into account a spacing between the scaled images in the overview image, a spacing of the scaled images from one or more of the boundaries of the overview image and / or other predefined conditions.

According to further exemplary embodiments, the pattern finder prepared in this way prepares for one or more local maxima in the census-transformed version of the overview image or in the version of the overview image transferred to the Hough feature space or in the gradient image Version of the overview image, or aigemein transferred to a Merkmaisraum version of the Übers tsbildes, identify a position of a local Ma ximums indicates the position of the identified predetermined pattern in the respective copy of the received and scaled image.

According to further embodiments, the pattern finder comprises a classification and a post-processing. The classification is applied to the overview image transferred to the mark space and provides high values (local maxima) at locations where the image content matches the searched pattern. Conventional methods according to the prior art can be used for the classification (eg, according to Viola-Jones).

The classified overview image can now be subjected to post-processing in accordance with further embodiments. For this purpose, the classified overview image is first smoothed with a local sum filter and locally corrected the positions of the local maxima in this way. Parallel to this, a local maximum filter is applied to the score (a score is in most cases the result of a classification, ie a measure of the match of the image content with the searched pattern) to the local maximum corrected via the cumulative filter receive. The result is again a classified overview image with Seores of the classifier but now locally corrected local maxima. Each position of a local maximum in the overview image is assigned a corresponding scaling level according to the previous embodiments (except at the points where the scaling levels are spaced apart from one another and to the boundaries of the overview image). Depending on exactly where the maximum lies within a scaling level, you can return this position to the original image by correcting the corresponding position by the scaling factor of the corresponding level of classification. Each position in the overview image is therefore assigned an absolute position in original image coordinates (except where the scaling levels are spaced from one another and to the boundaries of the overview image). If a maximum has now been extracted within a scaling stage, its position can also be corrected once more by using the corresponding local maximum from the adjacent scaling stages and averaging over the adjacent scaling stages.

Further embodiments relate to the precise implementation of the 21) image analyzer, which according to a first group of embodiments may be embodied as an embedded processor or programmable logic or as a custom unit (ASIC) or according to a second group of embodiments. Examples can be implemented as a method that runs on a computer (Compu ierprogramm).

Thus, one embodiment includes the method of analyzing a 2D image comprising the steps of: scaling a received image having a searched pattern according to a scale factor and producing an overview image having a plurality of copies of the received and scaled image, each Grouping is scaled by a different scaling factor. As a next step, the overview image is converted into a feature space (eg Hough feature space) and then a classification to determine a measure of the match with a predetermined pattern (such as an eye). Thereafter, search is made for maxima in the plurality of received and scaled images within the overview image to output information regarding a position at which a match between searched pattern and the predetermined pattern is maximum, the position being at a respective copy of the received and relates to a scaled image. If necessary, the position can be corrected using a combination of local sum and maximum filters and averaged over adjacent scaling steps.

According to further embodiments, the 2D bi-axis analyzer may be combined into a 2D image analysis system, the analysis system then having an evaluation device that observes and status the predetermined pattern and in particular the pupil (or, more generally, the eye region or eye of a user) determined. In this case, it is possible, in particular, to prevent a lack of reaction of the eye, such as e.g. an unopenable eye, e.g. due to a microsleep to detect.

According to further embodiments, the 2D image analyzer may be connected to a further processing means comprising a selectively adaptive data processor adapted to exchange a plurality of sets of data with the image analyzer and to process those data sets. In this case, the data records are processed in such a way that only plausible data records are passed on, whereby implausible parts of the data record are replaced by plausible parts.

According to further exemplary embodiments, the 2D image analyzer can also be connected to a 3D image analyzer, which determines an orientation of an object in space (eg, a viewing angle based on at least a first set of image data in combination with additional information) two main units, namely position calculating means for determining the position of the Pattern in three-dimensional space and an orientation calculator for

Determination of an axis by just this pattern.

According to further exemplary embodiments, the 2D image analyzer may also be connected to a Hough processor, which in turn is subdivided into the following two subunits: Pre-processor configured to receive a plurality of samples, each comprising one image and the image of the respective one To rotate and / or mirror samples and to output a plurality of versions of the image of the respective sample for each sample, Hough transform means adapted to detect a predetermined searched pattern in the plurality of samples based on the plurality of versions in which a characteristic of the Hough transformation device that is dependent on the searched pattern is adaptable. The processing means for postprocessing the results of the invention is adapted to analyze the detected patterns and to output a set of geometry parameters describing a position and / or a geometry of the pattern for each sample.

Embodiments of the present invention will be explained below with reference to the accompanying drawings. 1 is a schematic block diagram of a 2D image analyzer according to one

Embodiment;

2a is a schematic block diagram of a Hough processor with a

A processor and a Hough transformation device according to an embodiment;

2b shows a schematic block diagram of a pre-processor according to an embodiment; Fig. 2c is a schematic representation of FIough cores for the detection of straight lines (sections);

3a is a schematic block diagram of a possible implementation of a

Hough transformation device according to an exemplary embodiment;

3b shows a single cell of a delay matrix according to an embodiment; 4a-d a Schernatisch.es block diagram of another implementation of a Hough transformation device according to an embodiment;

5a shows a schematic block diagram of a stereoscopic camera arrangement with two image processors and a post-processing device, wherein each of the image processors has a Hough processor according to exemplary embodiments;

Fig. 5b shows an exemplary recording of an eye to illustrate a with the

Device from Fig. 5a feasible viewing angle detection and to explain the viewing angle detection in the monoscopic case;

Fig. 6-7 further illustrations for explaining additional embodiments or aspects; 8a-e are schematic representations of optical systems; and

9a-9i further illustrations for explaining background knowledge for the Hough transformation device.

Hereinafter, embodiments of the present invention will be explained in detail with reference to the drawings. It should be noted that the same elements are provided with the same reference numerals, so that the description of which can be applied to each other or interchangeable.

FIG. 1 shows a 2D image analyzer 700 with an image scaler 702, an image generator 704 and a pattern generator 706.

The image scaler 702 is configured to receive an image 710 (see Fig. 7b) having a searched pattern 71 1. In the image scaler 702, the received image 710 is now scaled with different scaling factors. In this case, for example, can be made a reduction, as shown below. In this reduction, based on the image 71 0, a plurality of copies of the received and scaled image 710 ', 710 ", 710'" to 710 "'''' arise, it being noted that the number of scaling stages is not limited is. This plurality of scaled images 710 'through 710 """''are assembled by the image generator 704 into an overview image 712, wherein the overview image 712 comprises the plurality of copies of the received and scaled image 710' through 710 ''''. Here, the respective position of a scaled image be calculated within the overview image by an algorithm that takes into account a spacing between the scaled images in the overview image, a spacing of the scaled images to one or more of the boundaries of the oversize image and / or other predefined conditions. In this overview image 712, the desired pattern 71 1 can now be detected by means of the pattern finder 706.

Thus, instead of applying the pattern recognition to several scaling stages in succession, the pattern recognition is applied only once to the overview image, whereby the respective scaling stage is "coded" as location information within the overview image Advantageous is this type of processing of individual scaling stages, in particular on FPGA architectures, since here The different scaled images would each have to be held in memory and processed separately and then the results would be merged, so that the overview image would be generated once and then processed in one step FPGA architectures are optimally utilized.

According to further exemplary embodiments, it would also be possible to transform this overview image 710 into a feature image 710a, a census transformation typically being used here. On the basis of this census-transformed image 710a, it would also be conceivable to perform a classification in order to determine the false-color image 710b (see Fig. 7c),

Optionally, according to further embodiments, a post-processing device (not shown) may additionally be provided in the 2D image analysis unit 700 (see FIG. 5a). The post-processing facility allows application of a combination of local sum and maximum filters to correct the positions of the local maxima in the classified overview image. Optionally, an averaging of the position of a maximum, which was found in one of the scaling stages, can also take place with the corresponding maxima from the adjacent scaling stages. Thus, an averaging takes place via adjacent scaling stages in the overview image.

Further exemplary embodiments or variants of this 2D image analyzer are explained with reference to FIG. 7a.

2a shows a Hough processor 100 with a pre-processor 102 and a Hough transformation device 104. The pre-processor 102 represents the first signal processing. tion level and is information coupled to the Hough transformer 1 04. The Hough translator 104 includes a delay filter 106 which may include at least one, but preferably a plurality, of delay elements 108a, 108b, 108c, 110a, 110b and 110c. The delay elements 108a to 108c and 10a to 10c of the delay filter 106 are typically arranged as a matrix, that is, in columns 108 and HO and lines a to c, and are signal-coupled with one another. According to the Ausführungsbeispie! 2a, at least one of the delay elements 108a to 108c or 110a to 110c has an adjustable delay time, here symbolized by the "+/-" symbol, for driving the delay elements 108a to 08c and 110a to 1.10c or to control the same, a separate control logic or an AnSteuerregister (not shown) may be provided.This control logic controls the delay time of the individual delay elements 108a to 108c and 1 10a to 1 10c via optional switchable elements 109a to 109c and The Hough transform means 104 may comprise an additional configuration register (not shown) for initially configuring the individual delay elements 108a-108c and 110a-110c.

The purpose of the pre-processor 102 is to prepare the individual samples 1 12a, 12b and 1 12c so that they can be efficiently processed by the Hough transformation device 104. To this end, the pre-processor 102 receives the bi-file (s) 12a, 12b, and 12c and performs preprocessing, e.g. in the form of a rotation and / or in the form of a reflection, in order to output the multiple versions (cf., FIGS. 12a and 31a ') to the Hougl transformation device 104. The output may be serial if the Hough transformer 104 has a Hough core 106, or parallel if more than one Hough core is provided. That So that depending on the implementation of the n versions of the image either completely parallel, semi-parallel (ie only partially parallel) or serially output and edited. The pre-processing in the pre-processor 102, which serves the purpose of detecting a plurality of similar patterns (rising and falling straight line) with a search pattern or a Hough core configuration, is explained below on the basis of the first sample 12a.

For example, this sample can be rotated, eg rotated 90 °, to obtain the rotated version 1 12a '. This process of rotation is provided with the reference numeral 1 14. In this case, the rotation can take place either by 90 °, but also by 180 ° or 270 ° or generally by 360 ° / n, it being noted that depending on the downstream Hough transformation (see Hough transformation device 104), it can be very efficient can perform only a 90 ° turn. In these sub-aspects, referring to FIG. 2b and 2c received. Furthermore, the image 1 12a can also be mirrored to obtain the mirrored version 1 12a "The process of mirroring is designated by the reference numeral 16 1. The mirror 16 corresponds to reading back the memory from the rear, both starting from the mirrored version 1 32a "as well as of the rotated version 1 12a ', a fourth version can be obtained by a rotated and mirrored version 1 12a'" by carrying out either the process 1 14 or 1 16. On the basis of the reflection 1 16, then two similar patterns (eg, right half-circle and half-circle opened to the left) are detected with the same Hough-core configuration as described below.

The Hough transform means 104 is adapted to provide, in the version 1 12a or 1 12a '(or 1 12a "or 1 12a"') provided by the preprocessor 102, a predetermined searched pattern, such as an ellipse or to detect a segment of an ellipse, a circle or segment of a circle, or a straight line or a trench segment. For this purpose, the Fiheranordnvvng is configured according to the sought predetermined pattern. Depending on the particular configuration, some of the delay elements 108a to 108c and 110a to 110c are activated or bypassed, respectively. As a result, upon application of a picture strip of the image under investigation 12a or 12a 'to the transformation means 104 by the delay elements 108a to 108c, some pixels are selectively delayed, which corresponds to a buffer memory, and others are forwarded directly to the next column 110. Depending on the image file 1 12a or 1 32a 'read in, or, to be exact, on the image structure of the applied line of the image 1 12a or 1, curved or oblique geometries are "straight-curved." 12a ', high column sums occur in one of the columns 108 or 110 while the column sums in other columns are lower The column sum is output via the spa sums output 108x or I Ox, optionally an addition element (not shown) for forming the Column sum can be provided per column 108 or 1. With a maximum of one of the column sums, it is possible to deduce the presence of a searched image structure or a segment of the searched image structure or at least the corresponding degree of agreement with the sought structure in that, for each processing step, the image strip is shifted by one pixel or by one column 108 or 110, such that In each processing step, it can be determined from an initial histogram whether one of the structures sought is detected or not, or whether the probability of the presence of the sought-after structure is correspondingly high. Expressed in other terms, this means that exceeding a threshold of the respective column sum of the column 108 or 110 indicates the detection of a segment of the searched image structure, each column 108 or 110 indicating a searched pattern Expression of a desired pattern (eg angle of a straight line or radius of a circle) is assigned. It should be noted at this point that for the respective structure, not only the respective delay element 1 10a, 1 10b and 1 10c of the respective line 1 10 is relevant, but in particular the previous delay elements 108a, 108b and 108c in combination with the following Verzögerun selemenien 1 10a, 1 10b and 1 10c, According to the prior art, such structures or activations of delay elements or bypass are predetermined from the outset.

By means of the variable delay elements 108a to 108c or 110a to 10c (delay elements), the desired characteristic (that is to say, for example, the radius or the rise) can be adapted during operation. Since the individual columns 108 and 110 are coupled to one another, the delay time of one of the delay elements 108a to 108c or 110a to 110c changes the overall filter characteristic of the filter 106. The flexible adaptation of the filter characteristic of the filter 106 of the Hough Transformer 104, it is possible to adapt the Transformationkcrn 106 during the passage, so that, for example, dynamic image content such for small and large pupils, with the same Hough core 106 can be detected and tracked. The exact implementation of how the delay time can be adjusted is discussed in FIG. 3c. In order to allow more flexibility to the Hough processor 100 or the transformation device 104, all delay elements c 108a, 108b, 108c, 110a, 110b and / or 310c (or at least one of the ones mentioned) are preferably provided with one variable or discretely schalibaren delay time, so that during operation active between different patterns to be detected or between different forms of the patterns to be detected can be switched over and over.

According to further embodiments, the size of the illustrated Hough core 104 is configurable (either in operation or in advance) so that additional Hough cells may be enabled or disabled.

According to further embodiments, the transformation means 104 may be provided with means for adjusting the same or, to be specific, for adjusting the individual delay elements 108a-108c and 110a-110c, such as with a controller (not shown). The controller is arranged, for example, in a downstream processing device and is designed to adapt the delay characteristic of the filter 106 if no pattern can be detected or if the recognition is not sufficiently good (low match of the image input). Folded with the searched patterns of existence of the searched pattern). This controller will be discussed with reference to FIG. 5a.

The abovementioned Ausführangsbei game has the advantage that this is easy and flexible to implement and in particular can be implemented on an FPGA. The background to this is that the parallel Hough transformation described above manages without regressions and is, as it were, completely parallelized. Therefore, further embodiments relate to FPGAs comprising at least the Hough transformer 104 and / or the pre-processor 102. In an implementation of the device described above on an FPGA, e.g. a XIL1NX Spartan 3A DSP, a very high frame rate of, for example, 60 FPS at a resolution of 640x480 could be achieved using a 96 MHz clock frequency because of the structure 104 described above with the plurality of columns 108 and 1 10 a parallel processing or a so-called parallel Hough transformation is possible.

2a and 2b show the pre-processor 102, which is used for pre-processing of the video data stream 1 12 with the frames 1 12a, 1 12b and 1 12c. The pre-processor 102 is designed to receive the samples 1 12 as binary edge images or also as gradient images and to carry out the rotation 114 or the reflection 1 16 on the basis of these in order to produce the four versions 1 12a, 12a ', 12a "and 1 12a '". The background to this is that typically the parallel Hough transform as performed by the Hougfa transform means is limited to two or four preprocessed, e.g. offset by 90 °, versions of a picture 1 12a builds. As shown in Fig. 2b, a 90 ° rotation (1 12a to 112a ') first takes place before the two versions 1 12a and 1 12a' are mirrored horizontally (compare 1 12a to 1 12a "and 1 12a ') 1 12a '"). In order to carry out the mirroring 16 and / or the rotation 14, the preprocessor, in corresponding exemplary embodiments, has an internal or external memory which serves to hold the received image files 12.

The processing rotation 14 and / or mirror 16 of the pre-processor 102 depends on the downstream FIough transformer, the number of parallel harmonics (degree of parallelization) and the configuration thereof, in particular as described with reference to FIG. 2c , In this respect, the pre-processor 102 may be configured to output the preprocessed video stream via the output 126 according to one of the three following constellations, depending on the degree of parallelization of the downstream Hough transcription unit 104: 100% Parallelization: Simultaneous output of four video streams, namely a non-rotated and non-mirrored version 1 12a, a 90 ° rotated version 1 12a ', and a mirrored version 1 12a "and 1 12a'" respectively. 50% parallelization; Output of two video data streams, namely non-rotated 1 12a and 90% mirrored 1 12a 'in a first step and output of the respective mirrored variants 1 12a "and 1 12a'" in a second step.

25% parallelization: each output of a video data stream, namely non-rotated 1 12a, rotated by 90 ° 1 12a ', mirrored 1 12a "and mirrored and rotates 1 12a'" sequentially.

As an alternative to the above variant, it would also be conceivable that, based on the first version, three further versions could be produced solely by rotation, that is to say by rotation through 90 °, for example.

180 ° and 270 °, based on which the Hough transformation occurs.

According to further embodiments, the pre-processor 102 may be configured to include other image processing steps, such as up-sampling. perform. In addition, it would also be possible for the pre-processor to generate the gradient image. ^If the gradient image generation becomes part of the image pre-processing, the gray value image (output image) in the FPGA could be rotated.

Figure 2c shows two Hough Kerukonilgurationen 128 and 130, eg for two parallel 3 1x3 Hough Heme, configured to recognize a straight or straight section. Furthermore, a unit circle 132 is plotted to illustrate in which angular ranges the detection is possible. It should be noted at this point that the hauer kernel gauges 128 and 130 are each to be seen so that the white dots illustrate the delay elements. The Hou gh core confi guration 128 corresponds to a so-called Type 1 Hough core, while the Hough core configuration 130 corresponds to a so-called Type 2 Hough core. As can be seen from the comparison of the two Hough core configurations 128 and 130, one represents the inverse of the other. With the first Hough core configuration 128, a straight line in the region 1 between 3π / 4 and π / 2 can be detected, while a straight line in the range 3π / 2 and 5π / 4 (range 2) is detectable by means of the Hough core configuration 130. In order to enable detection in the other areas, as described above, the Hough core configuration 128 and 130 is applied to the rotated version of the respective image. Consequently, by means of the Hough core configuration 128, the range lr between π / 4 and zero and, by means of the Hough core configuration 130, the range 2r between π and 3π / 4 can be detected. 2015/052009

Alternatively, using only one Hough core (e.g., a Type 1 Hough core), it may be useful to rotate the bit once by 90 °, once by 180 °, and once by 270 ° to detect the straight line orientation variations described above. On the other hand, due to the flexibility in the configuration of the Hough core, only one Hough core type can be used, which is reconfigured during operation or in which the individual delay elements can be switched on or off, that of the Hough core corresponds to the inverted type. In other words, this means that with the use of the pre-processor 102 (in 50% parallelization operation) and the configurable Hough transformation device 104 with only one Hough core and with only one image rotation, the complete functionality can be mapped otherwise, it can be understood that the particular Hough core configuration or the Hough core type choice is dependent on the pre-processing performed by the pre-processor 102 ,

Fig. 3a shows a Hough kernel 104 with m columns 108, 110, 138, 140, 141 and 143 and n rows a, b, c, d, e and f such that m x n rows are formed. The column 108, 110, 138, 140, 141 and 143 of the filter represents a particular feature of the structure sought, e.g. for a certain curvature or a certain straight rise.

Each cell comprises a delay element which can be set with respect to delay time, wherein in this exemplary embodiment the adjustment mechanism is realized by providing in each case a switchable delay element with a bypass. In the following, the structure of all cells will be explained by way of example with reference to FIG. 3b. The cell (108a) of Fig. 3b comprises the delay element 142, a remotely operable switch 144, such as e.g. a multiplexer, and a bypass 146. By means of the remote-controllable switch 144, either the line signal can be routed via the delay element 142 or fed to the node 148 without delay. The node 148 is connected on the one hand to the sum element 150 for the column (for example 108), on the other hand via this node 148 also the next cell (for example 1 10a) is connected.

The multiplexer 1 4 is configured via a so-called configuration register 160 (see Fig. 3a). It should be noted at this point that the reference numeral 160 shown here refers only to a portion of the configuration register 160 which is coupled directly to the multiplexer 144. The element of the configuration register 160 is configured to control the multiplexer 144 and receives, via a first information input 160a, configuration information that originates, for example, from a configuration matrix stored in the FPGA-internal BRAM 163. This conflagration tion information can be a column-wise bit string and refers to the configuration of several of the (also during the transformation) configurable delay cells (142 + 144). Therefore, the configuration information can be further forwarded via the output 160b. Since the reconfiguration is not possible at any time during operation, the configuration register 160 or the cell of the configuration register 160 receives a so-called Enabler signal via a further signal input 160c, by means of which the reconfiguration is initiated. The background to this is that the reconfiguration of the hauer requires a certain amount of time, which depends on the number of delay elements or, in particular, on the size of a column. In this case, one clock cycle is assigned for each column element and it is grain red to a latency of a few clock cycles by the BRAM 163 or the configuration logic 160. The overall latency for the reconfiguration is typically negligible for video-based image processing. It should be understood that in the present embodiment, the video data streams recorded with a CMOS sensor have horizontal and vertical blanking, and the horizontal blanking or horizontal blanking time can be used for reconfiguration. Through this relationship, the size of the Hough core structure implemented in the FPGA dictates the maximum size possible for Hough core configurations. For example, if smaller configurations are used, they are vertically centered and aligned horizontally on column 1 of the Hough core structure. Unused elements of the Hough core structure are all populated with activated delay elements.

The evaluation of the data streams thus processed with the individual delay cells (142 + 144) takes place column by column. For this purpose, the sum is added by columns in order to detect a local sum maximum, which indicates a recognized desired structure. The summation per column 108, 110, 138, 140, 141 and 143 serves to determine a value that is representative of the degree of agreement with the sought structure for an expression of the structure assigned to the respective column. For determining the local maxima of the column sums, 108, 110, 138, 140, 141 or 143 so-called comparators 108v, 110, 138v, 140v, 141v and 143v are provided per column, which are connected to the respective summation elements 150 , Optionally, between the individual comparators I08v, HOv, 138v. 140v, 141v, 143v of the different column 108, 1 10, 138, 140, 141 or 143 also further delay elements 153 may be provided, which serve to compare the column sums of adjacent columns. In detail, as the filter passes through, the column 108, 110, 138 or 140 is always passed out of the filter with the greatest degree of agreement for a characteristic of the desired pattern. Upon detection of a local maximum of a column sum (comparison previous, subsequent column), the presence of a searched structure can be concluded. be sen. The result of the comparison is thus a column number (possibly including the column sum = measure of agreement) in which the local maximum was recognized or the expression of the structure sought, eg column 138. In a preferred manner, the result comprises a so-called multidimensional Hough space, which contains all the relevant parameters of the desired structure, such as type of pattern (eg straight line or semicircle), degree of conformity of the pattern, extent of the structure (curvature of curve segments or slope and length of straight line segments) and Location or orientation of the searched pattern includes. In other words, for each point in the Hough space, the gray values of the corresponding structures in the image area are added up. As a result, maximums are formed by means of which the sought-after structure in Hough space can be easily localized and returned to the image area.

The Hough-core cell of FIG. 3b may include an optional pipeline delay element 162, arranged, for example, at the output of the cell and configured to receive both the signal delayed by means of the delay element 142 and that by means of the bypass 146 delay delayed signal.

As indicated with reference to FIG. 1, such a cell can also have a variable-area distribution element or can comprise a multiplicity of delay elements connected and bypassed with one another, so that the delay time can be set in several stages. In this respect, further implementations via the implementation of the Hough core cell shown in FIG. 3b would be conceivable as an alternative. An application of the device described above is explained below in FIG. 5a in an image processing system 1000. FIG. 5a shows an FPGA-implemented image processor 10a with a preprocessor 102 and a Hough transform device 104. Before the preprocessor 102, FIG Further, an input stage 12 may be implemented in the image processor 10a configured to receive image data or image sample from a camera 14a. For this purpose, the input stage 12 may comprise, for example, an image transfer interface 12a, a segmentation and edge detector 12b and means for camera control 12c. The camera control means 12c are connected to the image interface 12a and the camera 14 and serve to control factors such as gain and / or exposure.

The image processor 10a further includes a so-called Hough feature extractor 16 adapted to receive the multi-dimensional Hough space output by the Hough transform means 1.04 and all relevant information for pattern recognition involves analyzing and outputting a compilation of all Hough features based on the analysis result. In detail, a smoothing of the Hough feature spaces takes place here, ie a spatial smoothing by means of a local filter or a thinning out of the Hough space (suppression of non-relevant information for the pattern recognition). This thinning is done taking into account the nature of the pattern and the nature of the structure so that non-maxima in the Hough probability space are masked out. Furthermore, threshold values can also be defined for the thinning, so that, for example, minimum or maximum permissible characteristics of a structure, such as a minimum or maximum curvature or a smallest or largest increase, can be determined in advance. By means of this threshold-based suppression, noise suppression in the 1 lough probability space can also be carried out.

The analytic inverse transformation of the parameters of all remaining points in the original image area yields e.g. following Hough features: For the curved structure, position (x and y coordinates), probability of occurrence, radius and angle, which indicates in which direction the sheet is opened, can be forwarded. For a straight line, parameters such as position (x and y coordinates), probability of occurrence, angle indicating the slope of the straight line, and length of the representative straight section can be determined. This thinned-out Hough-Rautn is outputted by the FIough feature extractor 16 or generally by the image processor 10a to a post-processing device 18 for further processing.

This post-processing device can, for example, be implemented as an embedded processor and, depending on the application, have different subunits, which are explained below by way of example. The processing device 18 may include a Hough feature-to-geometry converter 202. This geometry converter 202 is configured to analyze one or more predefined searched patterns output by the Hough feature extractor and output the geometry describing parameters per sample. For example, the geometry converter 202 may be configured to output geometry parameters such as first diameter, second diameter, tilt and position of the center at an ellipse (pupil) or a circle based on the detected Hough features. According to a preferred embodiment, the geometry converter 202 is operable to detect and select a pupil based on 2 to 3 Hough features (eg, curvatures). Criteria, such as the degree of agreement with the sought-after structure or the Hough features, the curvature of the Hough features or the predetermined pattern to be detected, the location and orientation of the Hough features. The selected ones Hough feature combinations are sorted, where in the first line sorting is according to the number of Hough features obtained and secondarily after May3 matching the searched structure. After sorting, the first Hough-Fcature combination is selected to fit the ellipse that most closely represents the pupil in the camera's back.

Furthermore, the post-processing device 18 comprises an optional controller 204 which is adapted to output a control signal back to the image processor 10a (see control channel 206) or, in other words, back to the Hough transform means 104 on the basis of which Filter characteristic of the filter 106 is adjustable. For dynamic adaptation of the filter kernel 106, the controller 204 is typically coupled to the geometry converter 202 to analyze the geometry parameters of the detected geometry and to track boundaries defined by the Hough kernel i such that more accurate geometry recognition is possible. This process is a successive process that starts, for example, with the last flough kernel configuration (size of the last used Hough kernel) and is updated as soon as the recognition 202 yields insufficient results. Thus, in the example of the pupils discussed above, the controller 204 may adjust the ellipse size, e.g. is dependent on the distance between the object to be photographed and the camera 14a when the associated person nourishes the camera 14a. The filter characteristic is controlled on the basis of the last settings and on the basis of the geometry parameters of the ellipse.

According to further embodiments, the post-processing device 18 may comprise a selective-adaptive data processor 300. The purpose of the data processor is to rework outliers and dropouts within the data series in order to smooth the data series, for example. That's why the selective-adaptive

Data processor 300 is configured to receive a plurality of sets of values output by geometry converter 202, each set being assigned to a respective sarnple. The filter processor of data processor 300 performs a selection of values based on the multiple sets such that the data values of implausible sets (eg, outliers or dropouts) are replaced by internally-determined data values (substitute values) and the data values of the remainder Phrases continue to be used unchanged. In detail, the data values of plausible sentences (containing no outliers or suspects) are forwarded and the data values of implausible sentences (containing outliers or dropouts) are represented by data values of a plausible sentence, e.g. For example, the previous data value or an averaging of several previous data values replaced. The resulting data series of forwarded values and, if necessary, substitute values will be used smoothed out. This means that an adaptive temporal smoothing of the data series (eg a determined ellipse midpoint coordinate), e.g. B. on the principle of exponential

Smoothing. misfires or outliers (eg false detection in pupil detection) in the data series to be smoothed do not lead to fluctuations in the smoothed data. In detail, the data processor can smooth over the data value of a newly received sentence if it does not meet one of the following criteria:

According to the corresponding degree of agreement, which is provided by one of the

Additional value of the sentence is quantified, with the sought structure is a dropout in the data series.

According to the associated size parameter or geometry parameter is a dropout when z. For example, the size of the current object is too strong

Size of the previous object is different.

According to a comparison of the current data value with threshold values, which were determined based on the preceding data values, it is an outlier if the current data value (eg the current position value) is not between the threshold values. An illustrative example of this is when z. For example, the current position coordinate (data value of the sentence) of an object deviates too much from the ionization coordinate previously determined by the selectively adaptive data processor.

If one of these criteria is fulfilled, the previous value is still output or at least used to smooth the current value. In order to obtain the least possible delay in the smoothing, the current values are optionally weighted more heavily than the past values. For example, using exponential smoothing, the current value can be determined by the following formula:

Currently smoothed value = current value x smoothing coefficient + last smoothed value x (1 - smoothing coefficient)

The smoothing coefficient is dynamically adjusted within defined limits to the trend of the data to be smoothed, eg reduction in rather constant value curves or increase in ascending or descending value curves. If, in the long term, there is a major jump in the geometry parameters (ellipse parameters) to be smoothed, the data processor and thus also the smoothed value curve adapt to the new value. In principle, the selective-adaptive data processor 300 can also be configured by means of parameters, for example during the initialization, with the parameters behavior, x, b. maximum duration of a misfire, or maximum smoothing factor.

Thus, the selective adaptive data processor 300 or generally the post-processing device 118 can plausibly output values describing the position and geometry of a pattern to be recognized with high accuracy. For this purpose, the post-processing device to an interface 1 8a, via the optional control commands can be received externally. If several data series are to be smoothed, it is conceivable to use a separate selectively adaptive data processor for each data series or to adapt the selectively adaptive data processor so that data sets of different data series can be processed per set.

Hereinafter, the characteristics of the selective adaptive data processor 300 explained above with reference to a specific embodiment will be described generally:

The data processor 300 may, for. B. have two or more inputs and an output. One of the inputs (receives the data value and) is for the data series to be processed. The output is a smoothed row based on selected data. For selection, the further inputs (receive the additional values for a more accurate assessment of the data value) are used and / or the data series itself. During processing within the data processor 300, a change in the data rub takes place between the treatment of outliers and the handling of dropouts within the data series.

Outliers: In the selection, outliers (within the data series to be processed) are sorted out and replaced by other (internally determined) values.

Dropouts: One or more additional input signals (additional values) are used to assess the quality of the data series to be processed. The assessment is based on one or more threshold values, which subdivide the data into "high" and "low" grades. Low quality data is rated as a dropout and replaced by other (internally determined) values.

In the next step, for example, there is a smoothing of the data series (eg exponential smoothing of a time series). For smoothing, the data series adjusted by dropouts and outliers is used. The smoothing can be done by a variable (adaptive) coefficient. The G-factor is adjusted to the difference in the level of the data to be processed. According to further exemplary embodiments, it is also possible for the post-processing device 18 to have an image analyzer. such as a 3-bit analyzer 400. In the case of a 3D image analyzer 400, the post-processing device 18 can also be provided with a further image capture device consisting of image processor 10b and camera 14b. Thus, the two cameras 14a and 14b as well as the image processors 10a and 10b form a stereoscopic camera arrangement, wherein beforehand the image processor 10b is identical to the image processor 10a.

The 3D image analyzer 400 is configured to generate at least a first set of image data, which is determined on the basis of a first image (see camera 14a), and a second set of image data, which is based on a second image (cf. Camera 14b), wherein the first and second images map a pattern from different perspectives, and to calculate an angle of view and an SD view vector, respectively. To this end, the 3D image analyzer 400 includes a position calculator 404 and an alignment calculator 408. The position calculator 404 is configured to determine a position of the pattern in a three-dimensional space based on the first set, the second set, and a geometric relationship between the three Perspectives and the first and the second camera 14a and 14b to calculate. The alignment calculator 408 is configured to provide a 3D view vector, e.g. a viewing direction, according to which the recognized pattern is aligned in the three-dimensional space, the calculation being based on the first sentence, the second sentence and the calculated position (see position calculator 404). For this purpose, for example, a so-called 3 D amerasystemmodel 1 can be consulted, which has stored, for example, in a configuration file all model parameters, such as position parameters, optical parameters (see camera 14a and 14b).

The overall functionality of the 3D image analyzer 400 will now be described in detail on the basis of the example of the pupil recognition. The model stored or read in the 3D-Biidanalysator 400 includes data with regard to the camera unit, ie with respect to the camera sensor (eg pixel size, sensor size and resolution) and lenses used (eg focal length and ob] eke i nc), data or characteristics of the object to be recognized (eg, characteristics of an eye) and data relating to other relevant objects (eg, a display in the case of using the system 1000 as an input device). The 3D position calculator 404 calculates the Auger position or the pupil center point on the basis of the two or more camera images (cf. FIGS. 14a and 14b) by triangulation. For this purpose, he gets 2D coordinates of a point in the two camera images (see Figures 14a and 14b) made available via the process chain of image processors 10a and 10b, geometry converter 1er 202 and selective-adaptive data processor 300. From the transmitted 2D coordinates, the light beams which have imaged the 3D point as a 2D point on the sensor are calculated using the 3D camera model, taking into account in particular the optical parameters for both cameras 10a and 10b. The point of the two straight lines with the shortest distance to each other (ideally the intersection of the straight line) is assumed to be the position of the desired 3 D point. This 3D position, along with an error measure describing the accuracy of the passed 2D coordinates in conjunction with the model parameters, is either output via the interface 18a or passed to the view direction calculator 408.

Based on the position in the 3-D space, the viewing angle calculator 408 can determine the viewing direction from two elliptical projections of the pupil to the camera sensors without calibration and without knowledge of the distance between the eyes and the camera system. For this purpose, the viewing direction calculator 408 uses, in addition to the 3D positional parameters of the image sensors, the ellipse parameters which are determined by means of the geometry analyzer 202 and the position determined by means of the position calculator 404. From the 3D position of the pupil center and the position of the image sensors, virtual camera units are calculated by rotation of the real camera units whose optical axis extends through the 3D pupil center. Subsequently, projections of the pupil on the virtual sensors are respectively calculated from the projections of the pupil on the real sensors, so that two virtual ellipses are created. From the parameters of the virtual ellipse, the two sensors can each be calculated two points of view of the eye on any plane parallel to the respective virtual sensor plane. With the four viewpoints and the 3D pupil center points, four view direction vectors can be calculated, ie two vectors per camera. Of these four possible line of sight vectors, exactly one of the one camera is always approximately identical to one of the other cameras. The two identical vectors indicate the searched eye direction (gaze direction), which is then output from the sight line calculator 404 via the interface 1 8a.

A particular advantage in this 3D calculation lies in the fact that a non-contact and complete calibration-free determination of the 3D eye position and the pupil size independently of the knowledge of the position of the eye to the camera. rnera is possible. An analytical determination of the 3D eye position and SD viewing direction, including a 3D room model, allows any number of cameras (greater than 1) and any camera position in 3D space. The short latency with the simultaneously high frame rate enables a real-time capability of the described system 1000. Furthermore, the so-called time regimes are also fixed, so that the time differences between successive results are constant.

According to an alternative variant, it is also possible with only one camera also to perform a viewing direction determination, as explained below with reference to FIG. 5b.

In the previous description of the "3D image analyzer", which includes the method for calibration-free eye tracking, at least two camera images from different perspectives have been assumed so far.At the calculation of the viewing direction, there is a point at which exactly two possible pixels per camera are used In each case, the second vector corresponds to a mirroring of the first vector at the connecting line between the camera and the pupil center point .Thus of the two vectors which result from the other camera image, exactly one vector agrees with almost one calculated from the first camera image Vector match These matching vectors indicate the direction of view to be determined.

In order to be able to perform the calibration-free eye tracking also with a camera, of the two possible sight line vectors, hereinafter referred to as "vi" and "v2", which are determined from the camera wearer, the actual sight line vector (hereinafter referred to as "vb") must be used ) to be selected.

This process will be explained with reference to FIG. 5b. 5b shows the visible part of the eyeball (bordered in green) with the pupil and the two possible viewing directions vi and v2.

There are several ways to select the direction of view "vb", which can be used individually or in combination to select the actual line of sight vector, some of which (the list is not exhaustive) are explained below, assuming that vi and v2 (cf. FIG. 5a) have already been determined at the time of this selection:

According to a first possibility, one (the white dermis around the iris) can be done in the camera. There are 2 rays (starting at the pupillary center and finally long), one in the direction of vi and one in the direction of v2. The two beams are projected into the camera image of the eye and run there from the pupil center to the edge of the image. The ray, which passes over fewer pixels belonging to the sclera, belongs to the actual line of sight vector vb. The pixels of the sclera differ by their gray value from those of the iris bordering on them and those of the eyelids. This method reaches its limits if the face belonging to the recorded eye is too far away from the camera (ie the angle between the optical axis of the camera and the vector perpendicular to the face plane becomes too large).

According to a second possibility, an evaluation of the position of the pupil center can be made within the eye opening. The position of the pupil center within the visible part of the eyeball or within the eye opening can be used to select the actual line of sight vector. One way to do this is to define 2 rays (starting at the pupillary center and infinitely long), one in the direction of vi and one in the direction of v2. The two beams are projected into the camera image of the eye and run there from the pupil center to the edge of the image. The distance between the pupil center and the edge of the eye opening (shown in green in FIG. 5b) is determined along both beams in the camera image. The ray that results in the shorter distance belongs to the actual line of sight vector. This method reaches its limits if the face belonging to the recorded eye is too far away from the camera (ie the angle between the optical axis of the camera and the vector perpendicular to the face plane becomes too large).

According to a third possibility, an evaluation of the position of the pupil center can be made to a reference pupil center. The position of the pupum center point determined in the camera image within the visible part of the eyeball or within the eye opening can be used together with a reference pupil center to select the actual line of sight vector. One way to do this is to define 2 rays (starting at the pupillary center and infinitely long), one in the direction of vi and one in the direction of v2. The two beams are projected into the camera image of the eye and run there from the pupil center to the edge of the image. The reference pupil center within the eye opening corresponds to the pupil center when the eye is looking directly toward the camera sensor center of the camera used for imaging. The beam projected into the camera image, which has the smallest distance to the reference pupil center in the image, belongs to the actual sight line vector. To determine the reference pupil 11 There are several possibilities, some of which are described below:

Possibility 1 (special case of application): The center of the pupil center results from the determined pupil center, in the case where the eye looks directly towards the center of the camera sensor. This is the case if the pupil contour on the virtual sensor plane (see description for viewing direction calculation) describes a circle, option 2 (general application): The center of gravity of the area of the eye opening could be used as an estimate of the position of the reference pupil center point. This method of estimation reaches its limits when the plane in which the face lies is not parallel to the sensor plane of the camera. This constraint can be compensated if the inclination of the face plane to the camera sensor plane is known (eg, by a previously determined head position and orientation orientation) and this is used to correct the position of the estimated retraction pupil point.

Option 3 (general application): If the SD position of the center of the eye is available, a straight line between the 3D eye center and the virtual sensor center can be determined and its intersection with the surface of the eyeball. The reference pupil center results from the position of this intersection converted into the camera image. According to further exemplary embodiments, an ASIC (application-specific chip) can be used instead of the FPGAs 10a and 10b, which can be realized very cost-effectively, particularly in the case of high quantities. Overall, however, it can be stated that, regardless of the implementation of the Hough processor 10a and 10b, low energy consumption can be achieved by the very efficient processing and the associated low system clock demand.

Despite these features, the Ho ugh processor used here or executed on the Hough processor process remains very robust and störunanfällig. It is this

Note that the Hough processor 100 illustrated in FIG. 1 may be used in different combinations, in different combinations with features presented differently in particular with respect to FIG. 5. Applications of the H ou gh processor according to FIG. 1 are, for example, microsleepwatchers or fatigue detectors as driver assistance systems in the automotive sector (or in general at safety-relevant human machine interfaces), whereby evaluation of the eyes (eg covering of the pupil as a measure of the degree of opening) and taking into account the viewpoints and the focus, a specific pattern of fatigue can be detected. Furthermore, the Hough processor can be used on input devices or input interfaces for technical devices; Here the eye position and viewing direction are used as input parameters. A concrete application here would be the support of the user when viewing screen contents, eg when highlighting certain focused areas. Such applications are particularly interesting in the field of assisted iiving, in computer games, in the optimization of 3D visualization by including the line of vision, in market and media research or in ophthalmological diagnostics and therapies. As already indicated above, the implementation of the method presented above is platform-independent, so that the method presented above can also be executed on other units, such as a PC. Thus, another embodiment relates to a method for Hough processing comprising the steps of processing a plurality of samples, each having an image, using a pre-processor, wherein the image of the respective sample is rotated and / or mirrored, such that a plurality of versions of the image of the respective sample are output for each sample and detecting predetermined patterns in the plurality of samples based on the plurality of versions using a Hough transform means having a delay filter having a filter characteristic, the filter characteristic of which depends is set by the selected predetermined pattern set.

Although the above discussion has always referred to a filter characteristic in the context of adaptive characteristics, it should be noted that according to other embodiments, the adaptive characteristic may also relate to the post-processing characteristic (bending or distortion characteristic) for a fast correlation , This implementation will be explained with reference to FIGS. 4a to 4d. Fig. 4a shows a processing chain 1000 of a fast 2D correlation. The processing chain of the 2D correlation comprises at least the functional blocks 1 105 for 2D bending and 1 1 1 0 for the fusion. The procedure for 2D bending is illustrated in FIG. 4b. FIG. 4h shows examples of an overall position on templates. As based This processing chain 1000, a Hough feature can be extracted, will be apparent from Fig. 4c together with Fig. 4d. 4c shows the pixel-by-pixel correlation with n templates illustrated (in this case, for example, for straight lines of different pitch) for recognizing the ellipse 1 1 15, while FIG. 4d shows the result of the pixel-by-pixel correlation, typically with a maximum search via the n result images , Each result image contains a hough feature per pixel. Hough processing in the overall context is explained below.

In contrast to the implementation with a delay filter with an adaptable characteristic (implementation optimized for parallel FPGA structures), in the Hough processing shown here, which is predestined in particular for a PC-based conversion, part of the processing would be replaced by a different procedure It was the case that virtually every column of the delay filter represents a sought-after structure (eg straight line sections of different pitch). When passing through the filter, the column number with the highest sum value is decisive. The column number stands for a characteristic of the searched structure and the total value indicates a measure for the agreement with the searched structure.

In PC-based implementation, the narrative filter is replaced by fast 2D correlation. Depending on the size, the previous delay editor is able to map n characteristics of a specific pattern. These n values are stored as templates in the memory. Subsequently, the preprocessed image (z, B. binary edge image or gradient image) is traversed pixel by pixel. At each pixel position, all stored templates are compared with the underlying image content (corresponding to a postprocessing characteristic) (the environment of the pixel position (in terms of the size of the template, so to speak) is evaluated). This procedure is also referred to as correlation in digital image processing. For each template, therefore, one obtains a correlation value-that is, a measure of the match-with the underlying image content. These correspond, as it were, to the column sums from the previous delay filter. Now one decides (per pixel) for the template with the highest correlation value and remembers its template number (the template number describes the characteristics of the searched structure, eg slope of the straight line section).

So you get per pixel a correlation value and a template number. This allows a Houghfeature, as already explained, fully describe. It should also be noted that the correlation of the individual templates with the image content can be carried out both in the local and in the frequency domain. This means that the input image is first correlated with all n templates. One receives n resultants. If you put these result images on top of each other (like a cuboid) you would search per pixel for the highest correlation value (over all levels), where individual levels within the cuboid stand for individual templates. The result is again a single image, which then contains a correlation measure and a tempo number per pixel - that is, one hough feature per pixel. Although the above aspects have always been explained in the context of "pupil recognition", the aspects described above can also be used for other applications, such as the "microsleep" application, which will be discussed in more detail below. The second sleep detector is a system that consists of at least one of an image acquisition device, a lighting unit, a processing unit, and an acoustic and / or visual signaling device. By evaluating an image recorded by the user, the device is capable of detecting incoming microsleep or fatigue or distraction of the user and of warning the user.

This can be z. B. be configured in such a way that one uses a CMOS image sensor and illuminates the scene in the infrared range. This has the advantage that the device works independently of the ambient light and in particular does not dazzle the user. The processing unit used is an embedded processor system that executes a software code on an underlying operating system. The signaling device currently consists of a multi-frequency buzzer and an RGB LED.

The evaluation of the recorded image can take the form that in a first processing stage, face and eye detection and eye analysis are performed with a classifier. This level of processing provides initial clues for the orientation of the face, the eye positions, and the degree of lid closure.

Thawing on this, a model-based eye-contour analysis can be performed in the next step. An eye model used for this purpose can, for. Example, consist of: a pupil and / or iris position, a pupil and / or iris size, a description of the eyelids and the corner of the eyes. It is sufficient if at any time some of these components are found and evaluated. The one Components can also be tacked across multiple images so they do not have to be completely re-searched in each image.

The previously described Hough features can be used to perform face detection or eye detection or eye analysis or eye-contour analysis. The 2D image analyzer described above can be used for face detection or for eye detection or eye analysis. The described adaptive-selective data processor can be used for smoothing the result values or intermediate results or values profiles determined during face detection or eye detection or eye analysis or eye fine analysis.

A temporal evaluation of the LidscMussgtades and / or the results of the eye-fine analysis can be used to determine the microsleep or the tiredness or distraction of the user. In addition, the calibration-free sighting direction described in connection with the 1 ⁾ image analyzer can also be used to obtain better results in the determination of the microsleep or the fatigue or distraction of the user. To stabilize these results, the adaptive-selective data processor can also be used. The procedure for determining the eye position described in the exemplary embodiment "second scarf detector" can also be used to determine any other defined 2D position, such as a nose position or nose root position, in one

Be used face. When using a set of information from one image and another set of information, this position can also be determined in 3D space, the further set of information being from an image of another camera or by evaluating relations between objects in the first camera image can be generated. Accordingly, in one embodiment, the Hough processor in the image input stage may include camera control equipment.

Although some aspects have been described in the context of a device, it will be understood that these aspects also constitute a description of the corresponding method such that a block or device of a device is also to be understood as a corresponding method step or feature of a method step is.

Similarly, aspects described in connection with or as a method step also provide a description of a corresponding block or method Some or all of the method steps may be performed by an apparatus (using a hardware apparatus), such as a microprocessor, a programmable computer, or an electronic circuit. In some embodiments, some or more of the more important. Process steps are performed by such an apparatus.

Depending on specific implementation requirements, embodiments of the

Invention be implemented in hardware or in software. The implementation may be performed using a digital storage medium, such as a floppy disk, a DVD, a Blu-ray Disc, a CD, a ROM, a PROM, an EPRO, an EEPROM or FLASH memory, a hard disk, or other magnetic disk or optical Speiehers are stored on the electronically readable control signals that can cooperate with a programmable computer system or cooperate such that the respective method is performed. Therefore, the digital storage medium can be computer readable.

Thus, some embodiments according to the invention include a data carrier having electronically readable control signals capable of interacting with a programmable computer system such that one of the methods described herein is performed.

In general, embodiments of the present invention may be implemented as a computer program product having a program code, wherein the program code is operable to perform one of the methods when the computer program product runs on a computer.

The program code can also be stored, for example, on a machine-readable carrier. Other embodiments include the computer program for performing any of the methods described herein, wherein the computer program is stored on a machine-readable medium.

In other words, an exemplary embodiment of the method according to the invention is thus a computer program which has a program code for carrying out one of the methods described here when the computer program runs on a computer. A further embodiment of the invention is therefore like a process data carrier (or a digital storage medium or a computer-readable medium) on which the computer program for. Performing one of the methods described herein is recorded.

A further embodiment of the inventive method is thus a data stream or sequence of signals representing the computer program for performing any of the methods described herein. The data stream or the sequence of signals may be configured, for example, to be transferred via a data communication connection, for example via the Internet.

A further exemplary embodiment comprises a colorimg. For example, a computer or programmable logic device that is configured or adapted to perform one of the methods described herein.

Another embodiment includes a computer on which the computer program is installed to perform one of the methods described herein. Another embodiment according to the invention comprises a device or system adapted to transmit a computer program for performing at least one of the methods described herein to a receiver. The

Transmission can be done for example electronically or optically. The receiver may be, for example, a computer, a mobile device, a storage device or a similar device. For example, the device or system may include a file server for transmitting the computer program to the recipient.

In some embodiments, a programmable logic device (eg, a field programmable gate array, an FPGA) may be used to perform some or all of the functionality of the methods described herein. In some embodiments, a field programmable gate array may cooperate with a microprocessor to perform any of the methods described herein. In general, in some embodiments, the methods are performed by any hardware device. This may be a universally applicable hardware such as a Compiiterprozessor (CPU) or specific to the process hardware, such as an ASIC. In the following, the inventions described above or the aspects of the inventions will be described from two additional angles in other words:

Integrated Eyetracker

The "Integrated Eyetracker" includes a collection of FPGA-optimized algorithms that are capable of extracting (elliptical) features (Hough features) from a live camera image using a parallel flough transformation When using several cameras with known position and orientation, the 3D position of the pupil center as well as the 3D line of sight and the pupil diameter can be determined.The calculation uses the position and shape of the ellipses in the camera images.There is no calibration of the system required for the respective user and no knowledge about the distance between the cameras and the analyzed eye.

 The image processing algorithms used are in particular characterized in that they are optimized for processing on an FPGA (field programmable gate array). The algorithms enable very fast image processing with a constant frame rate, minimal latency and minimal resource consumption in the FPGA. Thus, these modules are predestined for time / latency / safety critical applications (eg driver assistance systems), medical diagnostic systems (eg perimeter) as well as applications such as human machine interfaces (eg for mobile devices) require low construction volume. problem

 Robust detection of 3D eye positions and direction in 3D space in several (live) camera images as well as detection of pupil sizes

 Very short reaction time (or processing time)

 Small construction form

 - autonomous operation (independence from the PC) through integrated solution

State of the art

 Eye Tracker Systems

 o Steffen Markert: Viewing the direction of the human eye in real time (diploma thesis and patent DE 10 2004 046 617 AI)

 Andrew T Duchowski; Eye Tracking Methodology: Theory and Practice Parallel-Hough Transformation

 o Johannes Katzmann; A real-time implementation for elliptical contour transformation (diploma thesis and patent DE 10 2005 047 160 B4) o Christian Holland-Neil: Implementation of a pupil

Detection algorithm based on the flough transformation for circles (diploma thesis and patent DE 10 2005 047 160 B4) Disadvantages of the current state of the art

 Eyetracker systems

 o Disadvantages:

 ■ Eyetracking systems generally require an (elaborate)

Calibration before use

^■ The system according to Markert (patent DE 0 2004 046 617 A1) is kaiib- rationsfrei, but only works under very specific conditions:

 i. Distance between cameras and pupil center must be known and stored in the system

 2, The procedure works only in case the 3D pupil center is in the optical axes of the cameras "All the processing is optimized for PC hardware and therefore also subject to its disadvantages (no fixed time regimen in the case of

Processing possible)

 "Powerful systems are needed because the algorithms have a very high resource consumption

 * long processing time and thus long delay time to result (z, T. depending on the image size to be evaluated)

Parallel-! iough transformation

 o Disadvantages:

^■ Only binary edge images can be transformed

^■ Transformation returns only a binary result related to an image coordinate (position of the structure was found, but not:

Hit probability and other structural features)

^■ No flexible adaptation of the transformation kernel during runtime and therefore insufficient suitability for dynamic image content (eg small and large pupils)

 ■ No reconfiguration of the transformation kernel to other structures during runtime possible, and thus limited suitability for object recognition

implementation

The overall system determines a list of multi-dimensional Hough features from two or more camera tabs in which the same eye is depicted, and calculates the position and shape of the pupillary ellipse on the basis thereof. From the parameters of these two ellipses as well as solely from the position and orientation of the cameras relative to one another, the 3D position of the pupillary midpoint as well as the 3D viewing direction and the pupil diameter can be determined completely without calibration. As a hardware platform we use a combination of at least two image sensors. FPGA and / or downstream. Microprocessor system (without a PC is absolutely necessary). "Hough preprocessing", "Parallel hough transform", "Hough feature extractor", "Hough feaiure to ellipse converter", "Coresize control", "Temporal smart smoothing filter ^" , "3D camera system model", "3D position calculation "and" 3D gaze direction calculation "relate to individual function modules of the integrated eye tracker, which are integrated in the image processing chain of the Integrated Eyetracker as follows:

6 shows a block diagram of the individual functional modes in the Integrated Eyctracker. The block diagram shows the individual processing stages of the Integrated Eyetracker. The following is a detailed description of the modules.

• "Hough preprocessing"

 o function

^■ Preparation of a video data stream for the module "Parallel

 Hough Transform ", in particular by image rotation and upsampling of the image to be transformed depending on the degree of parallelization of the module" Parallel Hough Transform "o Input

^■ Binary edge image or gradient image

 o Output

^■ Depending on the ^{degree of} parallelization of the subsequent module, one or more video data streams with prepared pixel data from the input

 o Detailed description

^{■ In} principle, the parallel Hough transformation can be applied to the image content from four main directions offset by 90 ° each

^{■ In} addition, the image is rotated by 90 ° during preprocessing

 ■ The two remaining directions are covered by mirroring the rotated and non-rotated image horizontally (by reading out the image matrix stored in the memory unequally)

 • Depending on the degree of parallelization of the module, the following three constellations result for the output:

 • 100% parallelization: simultaneous output of four video streams: rotated 90 °, not rotated and mirrored at a time

 • 50% parallelization: Output of two video data streams: rotated by 90 ° and not rotated, the output of the mirrored variants takes place sequentially

• 25% parallelization: Output of a video data stream: rotated by 90 ° and not rotated, and their respective mirrored variants are output sequentially "Parallel hough transform"

 o function

^■ Parallel detection of simple patterns (lines of different sizes and slopes and curves with different radii and orientations) and their probability of occurrence in a binary edge or gradient image

 input

 Edge or gradient image processed for the parallel Hough transformation (output of the "Hough preprocessing" module)

"Multidimensional Hough space containing all relevant parameters of the sought structure

 Detailed description

^■ Processing of the input through a complex delay-based local filter, which has a defined "pass direction" for pixel data and is characterized by the following properties:

 • Fi Variable-size headers consisting of delay elements

 • For adaptive adaptation of the filter to the desired patterns, delay elements can be switched on and off during runtime.

 • Each column of the filter stands for a certain characteristic of the sought structure (curvature or straight line increase)

• Aggregation over the column gives occurrence probabilities for the characteristic of the structure represented by the respective column

• When passing through the filter, the column with the highest probability of occurrence for a characteristic of the desired pattern is always passed out of the filter

^■ For each image pixel, the filter provides a point in the Hough space containing the following information;

 Type of pattern (e.g., straight or semicircle)

 • Occurrence probability for the pattern

• Formation of the structure (curvature of the curvature or in the case of straight lines: rise and length)

 • Location or orientation of the structure in the image

^■ The result of the transformation is a multi-dimensional image, which is referred to below as the "dream space""Houghfeatureextractor"

o function Extract features from the Hough space that contain relevant pattern recognition information

o fnput

 Multidimensional Hough Space (Output of the "Parallel Hough Transform" Module)

o Output

 "List of Hough features that contain relevant pattern recognition information

 Detailed description

 • Smoothing the Hough feature spaces (Spatial correction using local filtering)

 * "Thinning" of the Hough space (suppression of non-relevant information for the pattern recognition) by a modified "non-maximum-suppression":

 • Hiding points that are not relevant for further processing ("non-maxima" in the Hough probability space), taking into account the type of pattern and the structure's structure

 • Further thinning of the Houghraumpunkte with the help of suitable thresholds:

 o Noise suppression by threshold in the Hough

probability space

 o Specifying an interval for minimum and maximum permissible extent of the structure (for example, minimum / maximum curvature for curved structures or least / greatest increase for straight lines) ■ Analytical inverse transformation of the parameters of all remaining points into the original image area yields the following Hough features:

• Curved structures with the parameters:

 o Position (x and y image coordinates) o Occurrence probability of the Hough feature o Radius of the arc

 o Angle indicating in which direction the sheet is open

 • lines with the parameters:

o Position (x- and y-image coordinates) o Occurence probability of the Hough feature o Angle indicating the slope of the line o Length of the represented straight line section ugh fealure to eiUpse, Converter " o function

^■ Selecting the 3 to 4 Hough features (curvatures) most likely to describe the pupil margin (ellipse) in the camera image, and offsetting to an ellipse

 o Input

^■ List of all Hough features detected in a camera image

 (Curvatures)

 o Output

^■ Parameters of the ellipse that most likely represents the pupil

 o Detailed description

^■ From the list of all Hough features (curvatures), combinations of 3 to 4 Hough features are formed which, on the basis of their parameters, could describe the horizontal and vertical extreme points of an ellipse

^{■ The} following criteria are included in the selection of Hough features:

 • Scores (probabilities) of the Hough features

Curvature of the Hough features

 · Location and orientation of the Hough features to each other

^■ The selected Hough feature combinations are sorted:

• Primary based on the number of included Hough features

• Secondary according to the combined probability of the Hough features included

 ■ After sorting, the first ranked Hough

Feature combination selected and fitted out the ellipse, which most likely represents the pupil in the camera image

"Coresize contra L

 o function

^» Dynamic adaptation of the filter core (Hough-Core) of the parallel Hough transformation to the current ellipse size

 o Input

^■ Last used Hough-Core size

^■ Parameters of the ellipse representing the pupil in the corresponding camera image

 o Output

^■ Updated Hough-core size

 o Detailed description

^■ Depending on the size (length of the semiaxes) of the

"Hough feature to ellipse converter" calculated ellipse, the Hough core size is tracked within defined limits around the "EP2015 / 052009

accuracy of the Hough transformation results in the detection of elliptical extremes

"Tempora! smart smoothing senior "

 o function

^■ Adaptive temporal smoothing of a series of data (eg a determined ellipse center point coordinate) according to the principle of exponential smoothing, whereby dropouts or extreme outliers in the data series to be smoothed DO NOT result in fluctuations in the smoothed data

 o Input

^■ For each call-up time of the module, one value of the data series and the corresponding quality criteria (eg probability of occurrence of a gcfittetett ellipse)

 o Output

^■ Smoothed data value (eg ellipse center coordinate) o Detailed description

 "A set of filter parameters can be used when initializing the

Filters whose behavior is set

 * The current input value is used for smoothing if it does not fall into one of the following categories:

 • The corresponding probability of occurrence is a dropout in the data series

• The associated ellipse parameter is an outlier

 o If the size of the current ellipse is too different from the size of the previous ellipse

 o If the difference between the current position and the last position of the ellipse is too great

^■ If one of these criteria is fulfilled, the previously determined smoothed value will continue to be output, otherwise the current value will be used for smoothing

^■ To get the least possible delay in smoothing, current values are weighted more heavily than the past:

 • current smoothed value = current value * smoothing coefficient + last smoothed value * (1 - smoothing coefficient)

• The gating coefficient is dynamically adjusted within defined limits to the trend of the data to be smoothed: o Reduction in the case of a rather constant value progression of the data series

o Increase in ascending or descending value progression of the data series If, in the long term, a larger jump occurs in the ellipse parameters to be smoothed, the filter and thus also the smoothed value curve adapts to the new value

"3D camera system mode!"

 o function

^■ Modeling the 3 D room in which several cameras, the user (or his eyes) and possibly a screen are located o Input

^■ Configuration file containing the model parameters (position parameters, optical parameters, etc.) of all elements of the model o Output

^■ Provides a number skeleton and functions for calculations within this model

 o Detailed description

^■ Modeling of the spatial position (position and rotation angle) of all elements of the model as well as their geometric (eg pixel size, sensor size, resolution) and optical properties (eg focal length, lens distortion)

^■ The model includes the following elements at the moment:

• Camera units, consisting of:

 o camera sensors

 o lenses

 • Eyes

 • Display

^{■ In} addition to the properties of all elements of the model, in particular the following functions are provided: "3D position calculation" (for calculating the eye position) and "3D gaze correction calculation" (for calculating the viewing direction)

^■ This model can be used, among other things, to calculate the 3D line of sight (consisting of the center of the pupil and the directional vector (corrected according to the biology and physiology of the human eye))

^■ It is also possible to calculate the viewpoint of a viewer on another object in the 3D model (eg on a display) as well as the focused area of the viewer

,, 3D position calculation "

 o function

* Calculation of the spatial position (3D coordinates) of a point detected by two or more cameras (eg pupil center) by triangulation o Input

^■ 2D coordinates of a point in two camera images

o Output

^■ 3 D coordinates of the point

 "Feh! measure: describes the accuracy of the transferred 2D

Coordinates in connection with the model parameters

o Detailed description

 * Using the "3D camera system model" (in particular taking the optical parameters into account), the light beams that have depicted the 3D point as 2D points on the sensors are calculated from the transferred 2D coordinates for both cameras

^■ These rays of light are described as straight lines in the 3 D space of the model

^■ The point from which both straight lines have the smallest distance (ideally the intersection of the straight line) is assumed to be the position of the desired 3-point gaze direction calculation "

o function

 »Determination of the viewing direction from two elliptical projections of the pupil on the camera sensors without calibration and without knowledge of the distance between the eye and the camera system o Input

^■ 3 D position parameters of the brush sensors

 ■ Ellipse parameter of the pupil projected onto the two sensors

 • 3D positions of the ellipse centers on the two image sensors

^■ 3D position of the pupil center

o Output

 ■ 3D view in vector and world view

o Detailed description

^■ The 3D position of the pupil center and the position of the image sensors are used to calculate virtual camera units whose optical axis passes through the 3 D pupil center by rotating the real camera units

^■ Subsequently, from the projections of the pupil to the real sensors, projections of the pupil are calculated on the virtual sensors, so to speak two virtual ellipses are created

^■ From the parameters of the virtual ellipses two eye points of the eye can be calculated for each sensor on any plane parallel to the respective virtual sensor plane • Four viewpoints and the 3 D pupil center allow you to calculate four line of sight vectors (two vectors from each camera's results)

 * Of these four line of sight vectors, one camera is always (nearly) identical to the other camera

^■ The two identical vectors indicate the searched gaze direction, which is then returned by the "3D gaze direction cakulation" module. 4. a) Advantages

 Non-contact and completely calibration-free determination of the 3 Π-eye positions, 3 D-viewing direction and pupil size regardless of the knowledge about the position of the eye to the cameras

 Analytical determination of the 3D eye position and 3D viewing direction (including a 3-D space model) allows any number of cameras (> 2) and any camera positions in the 3D space

 Measurement of the pupil projected onto the camera and thus accurate determination of the pupil size

 - High frame rates (e.g., 60 FPS @ 640x480 on a X1L1NX Spartan 3A DSP @ 96MHz) and low latency through fully parallel processing without

Recursions in the processing chain

 Use of FPGA hardware and algorithms developed for the parallel FPGA structures

 Use of the Hough transformation (in the described adapted form for FPGA hardware) for robust feature extraction for object recognition (here:

Features of the pupil ellipse)

 Algorithms for postprocessing the Hough transformation results are optimized for parallel processing in FPGAs

 Fixed time regime (constant time difference between successive results)

 Minimal space, as fully integrated on chip

 Low energy consumption

 Possibility to directly port the processing from the FPGA to an ASIC ~> very cost-effective solution for large quantities by exploiting economies of scale

workarounds

 (partial) use of other algorithms (eg other object recognition method for elliptical detection)

Proof of patent infringement Apparently there could be a patient injury if the product in question is a fast and fully calibrated eyetracking system consisting of FPGA and microprocessor.

 Proof of patent infringement by copying / cloning the FPGA bitfile netlists

 o Would simply be detectable eg. By means of checksums

 o Furthermore, bitfiles can be bound to an FPGA ID -> copying would then only be possible if FPGAs with the same ID are used. Proof of patent infringement by "disassembling" the FPGA bit-n-bit network

 o Evidence for Patent Infringements could be identifiable by disassembling the relevant F P G A - B i part 1 / etsl o s The concrete evidence was difficult to provide

 Proof of patent infringement by "disassembling" the Prozxssor code

 o Clues would be apparent, concrete evidence would be difficult

application

 In a (live) camera image data stream, 3D eye positions and SD viewing directions are detected, which can be used for the following applications: o Safety-relevant areas

 "For example, microsleep detector or fatigue detector as a driver assistance system in the automotive sector, by evaluating the eyes (eg, covering the pupil as a measure of the degree of opening) and taking into account the viewpoints and the focus

 o human-machine interfaces

 * as input interfaces for technical devices (eye positions and viewing direction can be used as input parameters)

^■ Supporting the user when watching screen content (eg highlighting areas that are viewed)

^« Z. B.

 • in the field of Assisted Living

 • for computer games

 • Gaze-based input for Head. Mount Devices

• Optimization of 3D visualizations by incorporating the viewing direction

 o Market and media research

^{■ Evaluate,} for example, the attractiveness of advertising by evaluating the spatial direction of vision and the point of view of the test person o Ophthalmological diagnosis (eg objective perimetry) and therapy

1 PG Λ - 1 ^; ac etrac kc r One aspect of the invention relates to an autonomous (PC-independent) system, which in particular uses FPGA-optimized algorithms, and is suitable for this purpose. Detect face in a camera live image and determine its (spatial) position. The algorithms used are characterized in particular by the fact that they are optimized for processing on a FPGA (field programmable gate airay) and can do without processing recursions in comparison to the existing methods. The algorithms enable very fast image processing with a constant frame rate, minimal latency and minimal resource consumption in the FPGA. These modules are therefore predestined for time-critical applications (eg driver assistance systems) or applications such as human machine intcrfaces (eg for mobile devices), which require a low volume of construction. In addition, using a second camera, the spatial position of the user for specific points in the image can be determined with high precision, without calibration and without contact.

problem

 Robust and hardware-based face detection in a (live) camera image

 Detection of face and eye position in 3D space using a stereoscopic camera system

 Very short reaction time (or processing time)

 Small design

 autonomous operation (independence from the PC) through integrated solution

State of the art

 - Literature:

 o Christian Küblbeck, Andreas Ernst: Face detection and tracking in video sequences using the modified census transformation

 o Paul Viola, Michael Jones: Robust Real-time Object Detection Disadvantages of current Facetracker systems

 All processing is optimized for PC systems (general purpose processors) and thus also subject to their disadvantages (for example, no fixed time regime is possible during processing (example: tracking may take longer, depending on the image content, for example background))

 - Sequential processing; The input image is successively brought into different scaling levels (until the smallest scaling level is reached) and each searched with a multi-level classifier for faces

 o Depending on how many scaling levels have to be calculated or how many levels of the classifier have to be calculated, the processing time and thus the delay time will vary until the result is available

Powerful systems are needed to achieve high frame rates (higher clock rates, possibly multi-core systems), as they are already available on PC hardware optimized algorithms nevertheless have a very high resource consumption (especially in embedded processor systems)

 Based on the detected face position, classifiers only deliver inaccurate eye positions (the position of the eyes - in particular the pupil center - is not analytically determined (or measured) and is therefore high

Subject to inaccuracies)

 The determined face and eye positions are only available in 2D image coordinates, not in 3D

 The overall system determines the facial position from a cardiac image (in which a face is depicted) and, using this position, determines the positions of the pupil centers of the left and right eye. If two or more cameras with a known orientation are used, these two points can be specified for the 3-dimensional space. The two determined eye positions can be further processed in systems that use the "Integrated Eyetracker".

"Parallel image scaler", "Parallel face finder", "Parallel eye analyzer", "Parallel pupil analyzer", "Temporal smart smoothing filing", "1) camera system model" and "3D position calculation" relate to individual functional modules of the overall system (FPGA Facetracker) .They are placed in the image processing chain of the FPGA Facetracker as follows:

7a shows a block diagram of the individual function modules in the FPGA Facetracker 800. The function modules "3D camera system model 802" and "3D position calculation" 804 are not absolutely necessary for facetracking, but are made when using a stereoscopic camera system and offsetting suitable points used on both cameras to determine spatial positions (for example, to determine the 3-D head position when calculating the 2D facial centers in both camera images). The module "Feature extraction (Classification) 806" of the FPGA Facetracker builds on the feature extraction and classification of Küblbeck / Ernst from Fraunhofer IIS (Erlangen) and uses an adapted version of its classification based on Census characteristics.

The block diagram shows the individual processing stages of the FPGA Facetracking System. The following is a detailed description of the modules.

• "Parallel image scaler 702"

 o function

Parallel calculation of the scaling steps of the output image and arrangement of the calculated scaling steps in a new image matrix to enable the subsequent image processing modules to analyze all scaling stages simultaneously Figure 7b shows the output image 710 (original image) and result 712 (downscaling image) of the parallel image scaler. o Input

^■ Output image 710 in original resolution

 o Output 712

^■ New bus matrix containing several scaled variants of the output image in an arrangement suitable for the subsequent facetracking modules

 o Detailed description

^■ Structure of a picture pyramid by parallel calculation of the different scaling levels of the output picture

^■ In order to be able to ensure a defined arrangement of the previously calculated scaling stages in the target matrix, the image coordinates of the respective scaling stage are transformed into the image coordinate system of the target matrix on the basis of various criteria:

 • Defined minimum distance between the calibration stages in order to suppress crosstalk of analysis results into adjacent stages

 • Defines the distance to the edges of the target matrix in order to be able to partially guarantee the analysis of the image of outstanding faces

"Parallel face finds 808"

 o function

^■ Detects a face from classification results of several scaling levels, which are arranged together in a matrix. The parallel face finder 808 is similar to the finder 706 of FIG. 1, with the finder 706 including a generalized functionality (recognition of other patterns, such as pupil recognition). As shown in Fig. 7c, the result of the classification (right) represents the input for the parallel face fmder.

 o Input 712

^■ Classified image matrix containing multiple scaling levels of the source image

o Output ^• Position that is most likely to have a face (considering multiple criteria) o Detailed description

^■ Noise reduction to limit the classification results

^■ Spatial correction of the classification results within the scaling levels by means of a combination of local sum and maximum fiiter

 • Orientation to the highest probability of occurrence for a face, optionally at face size across all levels of skill

^■ Spatial averaging of result items across selected scaling levels

 • Selection of the scaling levels included in the averaging is made taking into account the following criteria: o Difference of the centers of the selected face in the considered scaling stages o Dynamic determined deviation from the highest result of the sum filter

 o Suppression of scaling stages without classification result

^■ Threshold-based setting of the detection performance of the "parallel face fmder"

"Parallel eye anaiyzer 81 0"

 o function

^■ During face detection, detects in parallel the position of the eyes in the corresponding face (this is especially important for faces that are not ideally frontal and twisted) o Input

^■ Image matrix containing several scaling levels of the output image (from the "Parallel image scaier" module) as well as the current position at which the most likely face is located (from the "Parallel face fmder" module) o Output

^■ Position of the eyes and a corresponding probability value in the face currently found by the "Parallel face fmder" o Detailed description

^■ Based on the scaled-down source image, the eye search described below for each eye is performed in a defined area (eye area) within the face region provided by the "Parallel Face Finder":

Defining the eye area from empirically determined normal positions of the eyes within the facial region With a specially shaped correlation-based local

Filters are detected within the eye area probabilities for the presence of an eye (the eye is described in this image area in simplified terms as a small dark area with a bright environment) The exact eye position including its probability results from a maximum search in the previously calculated probability complex rallel pupil analyzer 812 '

o function.

 * Detects the position of the pupillary centers within the detected eyes, based on a previously determined eye position (this increases the accuracy of the eye position, which is important for surveying or the subsequent evaluation of the pupil)

o Input

 * Original image in original resolution as well as the determined eye positions and face size (from the "Parallel eye analyzer" or "Parallel face finder")

o Output

 * Position of the pupil within the evaluated image as well as a status indicating whether a pupil was found or not. O Detailed description

 ■ Based on the determined eye positions and the face size, an image area to be edited around the eye is determined

^■ Across this image matrix, a vector is built that contains the minima of the image columns and a vector that contains the minima of the image lines

^■ Within these vectors (from minimal gray values), as shown in the following, the Pupü I enm titter unkt is detected separately in the horizontal and in the vertical direction:

 Detection of the minimum of the respective vector (as position within the pupil)

• Starting from this minimum within the vector, the position is determined in the positive and negative direction, at an adjustable percentage related to the ^'dynamic range of all vector elements threshold value is exceeded

The centers of these regions in the two vectors together form the center of the pupil in the analyzed image "Temporal smart smoothing filter 814"

 o function

^■ Adaptive temporal smoothing of a data series (eg, a determined face coordinate), where dropouts, nonsensical values or extreme outliers in the data series to be smoothed DO NOT result in fluctuations in the smoothed data

 o Input

^■ For each call-up time of the module, one value of the data series and the corresponding quality criteria (for face tracking: face score and downscaling level in which the face was found)

 o Output

^■ Smoothed data value (eg face coordinate)

 o Detailed description

^■ A set of filter parameters can be used when initializing the

Filters whose behavior is set

^■ The current input value is used for smoothing if it does not fall into one of the following categories:

 • The associated score is a dropout in the data series

 • The corresponding downscaling level is a nonsensical value (value found in a downscaling level that is too far away)

 • Excessive difference from the last value used for smoothing is an outlier

^■ If one of these criteria is fulfilled, the previously determined smoothed value will continue to be output, otherwise the current value will be used for smoothing

^■ To get the least possible delay in smoothing, current values are weighted more heavily than the past:

 • current smoothed value = current value * smoothing coefficient + last smoothed value * (1 - smoothing coefficient)

• The smoothing coefficient is dynamically adjusted within defined limits to the trend of the data to be smoothed: o Reduction in the case of more or less constant value progression of the data series

 o Increase in ascending or descending value progression of the data series

^■ If there is a long-term increase in the data to be smoothed or the associated downscaling levels, it will fit the filter and thus also the smoothed value curve to the new value

"3D camera system model 804a"

 o function

^■ Modeling the 3 D room in which several cameras, the user (or his eyes) and possibly a screen are located

input

 Configuration file containing the model parameters (attitude parameters, optical parameters, etc.) of all elements of the model

output

 * Provides a number skeleton and functions for calculations within this model

 Detailed description

⁸ Modeling of the spatial position (position and angle of rotation) of all elements of the model as well as their geometric (eg pixel size, sensor size, resolution) and optical (eg focal length, lens distortion) properties

^« The model includes the following elements at the present time:

• Camera units, consisting of:

 o camera sensors

 o lenses

 • Eyes

 • Display

 • In addition to the properties of all elements of the model, in particular the following functions are provided: "3D position calculation" (for calculating the eye position) and "3D gaze correction calculation" (for calculating the viewing direction)

^■ In other applications, the following functions are also available:

 • With the help of this model, u. a. the 3D gaze straight line (consisting of pupillary center and gaze vector (corrected according to the biology and physiology of the human eye))

 • Optionally, the viewpoint of a viewer on another object in the 3D model (eg, on a display) can be calculated as well as the focused area of the viewer

"3D position calculation 804"

o function ^■ Calculation of the spatial position (3D coordinates) of a point captured by two or more cameras (eg, pupil center)

 o Input

 ■ 2D coordinates of a point in two amera images

 o Output

^■ 3D coordinates of the point

^{■ Error} measure: describes the accuracy of the passed 2D coordinates in conjunction with the model parameters

 o Detailed description

^■ Using the "3D eamera system model" (especially taking into account the optical parameters), the 2D beams are used to calculate the light beams for both cameras that have reproduced the 3D point as 2i) points on the sensors Light rays are described as straight lines in the 3 D of the model

^■ The point from which both straight lines have the shortest distance (ideally the intersection of the straight line) is assumed to be the position of the desired 3 D point

advantages

 Determining the face position and the eye positions in a (live) camera image in 2D and by recalculating in the 3D space in 3D (by incorporating a 3 D space model)

 - The algorithms presented under 3 are optimized for real-time capable and parallel processing in FPGAs

 - High frame rates (60 FPS @ 640x480 on a XILINX Spartan 3A DSP @ 48 MHz) and short latencies due to fully parallel processing with no recursions in the processing chain -> very fast image processing and low-delay output of the results

 Minimal installation space, since all functionality can be achieved with one component (FPGA)

 Low energy consumption

 Fixed time regime (constant time difference between successive results) and thus predestined for use in safety critical applications

 Possibility to directly port the processing from the FPGA to an application specific integrated control (ASIC) -> very cost-effective solution in case of large quantities by exploiting economies of scale

workarounds Use of other algorithms for the Gcsarotfunktionaiität or individual sub-functions

Proof of patent infringement

 - Proof of patent infringement by copying / cloning the FPGA

Bitfile / netlists

 o Would simply be detectable eg. By means of checksums

 o Furthermore, bitfiles can be bound to an FPGA-iD - the data would only be copied if FPGAs with the same ID were used

 - Proof of patent infringement by "disassembling" the FPGA

Bitfile / netlists

 o Evidence of patent infringement might be identifiable by disassembling the relevant FPG A-Biteii 1 / N etting List o It would be difficult to provide concrete evidence

application

 Advantages in the application compared to the software solution

 o autonomous mode of operation (system on chip)

 o Possibility of easy transfer to an ASIC

 o Space-saving integration into existing systems / circuits

 Areas of application similar to those of a software solution (in a (live) camera data stream, face positives and the corresponding eye positions are detected, which can be used for the applications listed below)

 o Security applications

^{■ For} example, second-level detector in the automotive sector, by evaluating the eyes (opening degree) and the eye and head movement o Human-machine communication

^■ eg input interfaces for technical devices (head or eye position as input parameters)

 o gauze tracking

⁸ eg facial and eye positions as a precursor to sight direction determination (in combination with "Integrated Eyetracker") o Marketing

 ■ e.g. Assess the attractiveness of advertising by evaluating

Head and eye parameters (inter alia position)

In the following, further background knowledge on the above-described aspects of the inventions will be disclosed on the basis of two representations. A detailed calculation example for this view direction calculation will be described below with reference to Figs. 8a to 8e.

Calculation of the point of interest.

As already described, when the circular pupil 806a is imaged by the camera lenses 808a and 808b, an elliptical pupil projection is produced on the image sensors 802a and 802b (see FIG. The center of the pupil is always imaged on both sensors 802a and 802b and thus also in the corresponding camera images as the center EP ^KS and E _M p ^{K2 of} the ellipse. Dalier can be determined by stereoscopic backprojection of these two Ellipsenrnittelpunkte E _M p ^KI and EP ^K2 using the lens model of the 3D pupil center. An optional prerequisite for this is an ideally time-synchronized image acquisition so that the scenes reproduced by both cameras are identical and thus the pupil center has been recorded at the same position.

First, for each camera, the back projection beam RS of the ellipse center has to be calculated, which runs along the node beam between the object and the object-side node (Hl) of the optical system (FIG. 8a),

RS (l) = RSo + t - RS _s

 (A I)

This back projection beam is defined by equation (AI). It consists of a starting point RSo and a normalized directional vector RS-, which are used in the

Lens model (Figure 8b) by the equations (A2) and (A3) from the two main points H _\ and Hi of the objective and the ellipse center EMP in the sensor plane result. For this, all three points (Hj, H ₂ and EMP) must be present in the Eyetracker coordinate system.

RSo = Hi

(AT)

 (A3)

The main points can be with the equations H, and H, = K ₀ + (b + d) - K; calculate directly from the lens and camera parameters (Figure 8b), Ko is the center of the camera sensor plane and the normal vector is the camera sensor plane. The 3D ellipse center point in the camera coordinate system can be determined from the previously determined E-hopping center line parameters Tri x ", and y", which are present in image coordinates by means of an equation

Camera ^■ 5 ^* PxCr

calculate. Here P _B üd is the resolution of the camera image in pixels, S ₀ f _{fSe (} is the position on the sensor at which the image is read out, S _rcs is the resolution of the sensor and S xör is the pixel size of the sensor.

The desired pupil center is ideally the intersection of the two back propagation beams RS ^K1 and RS ^K2 . However, with practically determined model parameters and ellipse centers, no intersection of the straight lines in the 3-D space results, even with minimal measurement errors. Two straight lines in this constellation, which neither intersect nor run parallel, are called skewed straight lines in geometry. In the case of backprojection, it can be assumed that the two skewed straight lines each pass very close to the pupil center track. In this case, the pupil center is at the point of their smallest distance to each other halfway between the two lines.

The shortest distance between two skewed straight lines is indicated by a connecting line perpendicular to both straight lines. The direction _vector n _Sl

The distance perpendicular to both rear projection beams can be calculated according to equation (A4) as a cross product of their direction vectors.

(A4) The location of the shortest link between the back projection beams is defined by Equation (A5). By inserting RS ^kl (s), RS ^K2 (f), this results in a system of equations from which s, l and u can be calculated. RS ^KI (x) + un _SI RS

 (A5)

The sought pupil center P P, which lies halfway between the back projection beams, thus results from equation (A6) after the onset of the values calculated for s and u.

(A6) In addition, as an indicator of the accuracy of the calculated pupil center PMP, the minimum distance d _RS between the back projection beams can be calculated. The more accurate the model parameters and ellipse centers were, which were used for the previous computations, the smaller the d. d _{m ^} ~ \ n _l \

(A7)

The calculated pupil center is one of the two parameters that determine the eye's eye's line of sight. It is also needed to calculate the line of sight vector, which is described below.

The advantage of this method for calculating the pupillary center is that the distances between the cameras and the eye need not be stored permanently in the system. This is z. Example, in the described in the patent DE 1 0 2004 046 617 A I method required.

Calculation of the line of sight vector The sight line vector P _fi to be determined corresponds to the normal vector of the circular pupil surface and is thus determined by the orientation of the pupil in the 3-D space. From the ellipse parameters, which can be determined for each of the two elliptical projections of the pupil on the camera sensors, the position can be and orientation of the pupil. Here, the lengths of the two half-axes and the rotation angle of the projected ellipses are characteristic of the orientation of the pupil or the viewing direction relative to the camera positions. An approach for calculating the viewing direction from the ellipse parameters and fixed distances in the eye tracking system between the cameras and the eye is z. As described in the patent DE 10 2004 046 617 AI. As shown in FIG. 8e, this approach is based on a parallel projection, wherein the straight line defined by the sensor normal and the center point of the pupil projected onto the sensor runs through the pupil center. For this purpose, the distances between the cameras and the eye must be known in advance and stored firmly in the eye-tracking system.

By contrast, the model of the camera lens used in the approach presented here, which describes the imaging behavior of a real objective, is replaced by a perspective projection of the object onto the image sensor. As a result, the calculation of the pupil center can be made and the distances of the cameras to the eye need not be known in advance, which is one of the significant innovations over the above-mentioned patent. By perspective projection, however, the shape of the pupil ellipse imaged on the sensor does not result from the inclination of the pupil with respect to the sensor surface, unlike the image elimination. The deflection δ of the pupil center from the optical axis of the camera lens has, as shown in FIG 8b also shows an influence on the shape of the pupil 1 projection and thus on the ellipse parameters determined therefrom. In contrast to the sketch in Fig. 8b, the distance between pupil and camera with several hundred millimeters is very large compared to the pupil radius, which is between 2 mm. and 8 mm. Therefore, the deviation of the pupil 1projection from an ideal ellipse shape, which arises at a tilt of the pupil with respect to the optical axis, becomes very small and can be neglected.

In order to be able to calculate the line of sight vector P _n -, the influence of the angle δ on the ellipse parameters need not be eliminated, so that the shape of the pupil projection is influenced solely by the orientation of the pupil. This is always the case when the pupil center PMP lies directly in the optical axis of the camera system. Therefore, the influence of the angle δ can be eliminated by calculating the pupil projection on the sensor of a virtual camera system vK whose optical axis passes directly through the previously calculated pupil center MP, as shown in Fig. 8c. The position and orientation of such a virtual camera system 804a '(vk in Fig. 8c) can be calculated from the parameters of the original camera system 804a (K in Fig. 8b) by rotation about its object-side main point Ah. At the same time, this corresponds to the object-side main point v / Y _{/ of} the virtual camera system 804a '. Thus, the direction vectors of the node beams of the imaged objects in front of and behind the virtual optical system 808c 'are identical to those in the original camera system. All further calculations for determining the line of sight vector are carried out in the Ey etrac ker coordinate system.

The normalized normal vector vK "of the virtual camera vK is as follows:

 (A8)

For the further procedure it is necessary to calculate the rotation angles around the x-axis (νΚ _θ ), around the y-axis (νΚ _φ ) and around the z-axis (vK) of the eyetracker coordinate system by which the unit vector of the z Direction of the eye tracker coordinate system must be rotated about different axes of the eye tracker coordinate system to obtain the vector vK-. By rotation of the unit vector of the x-direction and the

Unit vector of the y-direction of the eye tracker coordinate system by the angles νΚ _θ , νΚ _φ and νΚ _ψ can the vectors VÄ ^' . and vK _y - which are the x- and y-

Specify the axis of the virtual sensor in the eye tracer coordinate system.

In order to obtain the position of the virtual camera system 804a '(Fig. 8c), its location vector or coordinate origin vKo, which is also the center of the image sensor, must be calculated by Equation (A9) to be in the node beam of the pupil center PM _P , AL = v / i, - (d + b vK,

 (A9)

The required distance d between the main points and the distance b between the main plane 2 and the sensor plane must be known or z. B. be determined experimentally with a test setup.

Furthermore, the position of the image-side main point from equation (Λ10). v // ₂ - vll .... d- vK _t

 (AIO)

To calculate the pupil projection on the virtual sensor 804a ', first edge points RP.sub.y ^{,} of} the previously determined ellipse on the sensor in the original position are required.

These arise at the. Edge points K ^{20 of} the ellipse E in the camera image, corresponding to Fig.8d E _A is .Short half-axis of the ellipse, E _{h is} the long semiaxis of the ellipse, E _SM and Ε _{ν ^} the center coordinates of the ellipse and E _{A of} the rotation angle of the ellipse. The position of a point RP '° in the eyetracker coordinate system can be calculated by the equations (AI 1) to (AI 4) from the parameters of the ellipse E, the sensor 5 "and the camera K, where ω is the position of a boundary point RP ^2D according to Fig.8d indicates on the EHipsenumfang.

(All) x'-cs {E + y'-sm (E _a ) + E ^

RP ²ⁿ =

 (A13)

RP ^w = K ₀ + s K _x . + T K,

 (AI 4)

The direction of a node beam KS in the original camera system, which images a pupil edge point as ellipse edge point RP ^{1 "} " on the sensor, is the same

Direction of the node beam VCS in the virtual camera system, which images the same pupil edge point as ellipse edge point RP ^{: s} on the virtual sensor. The nodal point rays of the ellipse edge points in Fig. 8b and Fig. 8c illustrate this. Thus, the two beams KS and vKS have the same directional vector, which results from equation (AI 5). For the position vector vKS of the virtual sensor-side node beam vKS, VKSQ = vEI ₂ always applies.

RP ⁱⁿ -ll,

RP ^iD - H, \ (WHEN)

The virtual node point beam and the virtual sensor plane corresponding to the x-y plane of the virtual camera vK are set equal in equation (A 16), where the parameters of their intersection result by resolving to $ z and h. With these, by equation (AI 7), the ellipse edge point can be calculated in pixel coordinates in the image of the virtual camera.

vKS "+ r ₇ - vKS

(AI 6)

 (A17)

Subsequently, from a plurality of virtual boundary points vRP ^{2 n} mitteis Ellipsenfitting, z. For example, with the Fitzgibbon et ah "Direct ieast square attaining ellipses" algorithm, the parameters of the virtual ellipse vE shown in Figure 8c may be computed by using at least six virtual boundary points vRP ²⁰ , by substituting different ω into equation (Al l) can be calculated by the above-described route.

The shape of the virtual ellipse vE thus determined depends only on the orientation of the pupil. Moreover, its center is always in the center of the virtual sensor and, together with the sensor normal, which corresponds to the camera normal vK-, forms a straight line running along the optical axis through the pupil center Ρμρ. Thus, the prerequisites are fulfilled in order to subsequently calculate the line of sight based on the approach presented in the patent DE 10 2004 046 617 A I. With this approach it is now also possible to determine the viewing direction by using the virtual camera system described above, when the pupil center is located outside the optical axis of the real camera system, which is almost always the case in real applications.

As shown in FIG. 8e, the previously calculated virtual ellipse vE is now assumed in the main virtual plane 1. Since the center of vE is at the center of the virtual sensor and thus in the optical axis, the 3-D ellipse center point VE 'MP corresponds to the virtual main point 1. It is also the lotus point of the pupillary center PMP in the main virtual plane 1. At the same time, only the axis sen proportions and the Roiationswinkel the ellipse ve used. This Fotmparaiiieter of vE can also be used unchanged with respect to the main level 1, since the alignments of the x- and y-axis of the 2-D sensor plane to which they relate, the orientation of the 3-D-Scnsorebene correspond and thus also the orientation of the main level 1.

Each image of the pupil 806a in a camera image can be formed by two different orientations of the pupil. In the evaluation of the pupil shape, therefore, as shown in FIG. He, two virtual intersections v S of the two possible straight lines with the main virtual plane 1 result from the results of each camera. According to the geometric V in Fig. He, the two possible viewing directions p. , and p. ₂ can be determined as follows.

The distance A between the known pupil center and the ellipse center vE 'MP is:

A = \ vH - PM _> P

(A18)

From this it is possible to determine r with equation AI 9.

(A19)

The two Richtarigsvektoren r ~, and _Ta , differs from vH,> on vSj and vS? are aligned, are analogous to the equations

0 0

 M Ψ cos (^) - sin (»)

 sin (<p) cos (ö) 0 sin (0)

 M Θ 0 1 0

- ύι \ (θ) 0 cos (ß)

Μ _β - Μ · Μ _Κ

calculated from VKQ, vK <, vK and vEV:

(A20)

[i, o, o

 (A21)

Afterwards, the two virtual intersections vS / and S _? and from this the possible directions of view p. , as well as t. , be determined. vS, = vH, + r ^■ r-

(A22) vS, = vH, + r · r _s

 (A23)

vS ^ P _h

 (A24) vS, - P,

vS ₂ - P,

(A25) In order to determine the actual viewing direction, the possible viewing directions of the camera 1 (p ^ sovk) around ^the camera 2 (j ^ and _h ^K f) are required. Of these four vectors, one from each camera indicates the actual viewing direction, and these two normalized vectors are ideally identical. In order to identify them, for all four possible combinations of one vector of one camera and one vector of the other camera, the differences of the respective selected sight line vectors are selected. The combination which gives the least difference contains the vectors sought. These give the mean to be determined In the case of averaging, an almost simultaneous image acquisition must be assumed, so that the same pupil position as well as the same orientation and thus the same view glare were captured by both cameras,

In addition, as a measure of the accuracy of the calculated line of sight _vector, the angle w _dijj between the two averaged vectors P ^KI and °, which indicate the actual viewing direction, can be calculated. The smaller M> dijf, the more accurate were the model parameters and ellipse centers used for previous calculations.

 (A26)

The angles of view 9 _B w and fair with respect to the normal position of the pupil (P ~ lies paraiiel to the z-axis of the eye tracker coordinate system) can be calculated using the equations

Β _ιν = and

be calculated.

If a systematic deviation of the viewing direction from the optical. Axis of the eye or is to be considered by the PupiUnormale, the corresponding angle to the determined viewing angles 0BW and ΨΒ can be added. The new view vector must then be given by means of equation from the new angles B _B W and ψβιν 'and 5 = [0,0, l} ^r .

With the line of sight vector Ρ, (next to the pupillary center PM _P of equation

A6) is also the second parameter known from the 3D image analyzer to be determined eye line (LoS). This results according to the following equation LoS t) = P _MP + l - P "-.

The implementation of the above-presented method is platform independent, so the above-presented method can be used on various hardware platforms, e.g. a PC can be executed.

Development of a method for the revision of the method characteristic of sextraction

The aim of the present following is to develop on the basis of the parallel Hough transformation and a robust method for feature extraction. This will be done by revising the Houghcore and introducing a feature extraction method that reduces the results of the transformation and breaks them down to a few "feature vectors" per image, then implements and tests the newly developed method in a Matlab toolbox, eventually becoming an FPGA implementation of the new procedure.

Parallel Hough transformation for straight lines and circles

The Hough parallel transformation uses Houghcores of different sizes, which must be configured using configuration matrices for each application. The mathematical relationships and methods for creating such configuration matrices are shown below. The Matlab alc config lines curvatures.m script uses these techniques to create straight-line and semicircle configuration matrices of various sizes. To create the configuration matrices, it is first necessary to calculate a set of curves in discrete representation and for different Houghcore sizes. The requirements (educational regulations) on the group of curves have already been shown. Under In particular, straight lines and semicircles are suitable for configuring the Houghcores, taking these educational regulations into account. For the viewing direction determination

Houghcores with configurations for semicircles (or, curvatures) used. For the sake of completeness, the configurations for straight lines (or straight sections) are also derived here. The mathematical relationships for determining the curves for straight lines are illustrated.

The starting point for the calculation of the curves for straight lines b is the linear straight line equation in (B l). y = m ^■ x + n

The families of curves can be generated by varying the rise m. For this, the straight line slope from 0 ° to 45 ° is split into equal intervals. The number of intervals depends on the Houghcore size and corresponds to the number of Houghcore

Zeiien. The increase can be tuned by the variable Y _core from 0 to core _he j _gl .

 (B2)

The function values of the curves are calculated by varying the variables (replaced by x _corc in (B3)) whose values run from 0 to core width.

¹

J ¥ ^" e.orc ^ ^'' core

core _hot!

 (B3)

For a discrete representation in the 2D plot, the function values must be rounded to integers. The calculation of the curves for semicircles is based on (Katzmann 2005, pp 37-38) and is illustrated in Fig. 9b.

The starting point for calculating the sets of curves is the circle equation in coordinate form, r ² = (x ^{^} X _M f + y - y _M f (B4)

With x _M = 0 (position of the center point on the y-axis), x = x _core and change to y follows for the function values of the sets of curves (B5).

 (B5)

Since and r are not known, they must be replaced. These can be the mathematical

Relationships in (B6) and (B7) are derived from Fig. 9b. y _M = h - r

(B6)

 (B7)

By changing from (B7) to y _M and the condition that y always has to be negative (s, Fig. 9b) one obtains (B8).

 (B8)

Insertion of (B8) into (B5) leads to (B9).

(B9)

From Fig. 9b it is clear that the Houghcore is center-centered in the y-axis of the circular coordinate system. The variable x _core normally runs from 0 to cofe _wjil! sj - 1 and must therefore be corrected by - · -.

 (B O)

The radius is missing, which is obtained by inserting (B6) in (B7) and by further transformations.

(A - r) ' ₊

(Bl l)

(B 12)

 (B 13) core hcighl

Finally, to generate the sets of curves, the variable h must be from 0 to

be varied. This is done via the variable y _core which runs from 0 to core, _h

r =

(B14)

As with the straight lines, the y-values for a discrete representation in the 2D plot must be rounded. The curves for Houghcores type 2 can be easily determined by equation (B 15). y ₂ = core _heifil - y.

(B15)

Starting from the curves, two configurations (type 1 and for straight lines and circles) can be determined for all Houghcore sizes. Ones are thereby determined directly from the group of curves, cf. (Katzmann 2005, pp. 35-36). Confi guration matrices can be either zeros or ones. A one stands for a used delay element in the Houghcore. First, the configuration matrix in the dimensions of the Houghcore is initialized to zero values. Then the following steps will be performed:

1 . Start with the first curve of the set of curves and check the y-value of the first x-index number. If the y value is greater than zero, then occupy the same line

(same y-index) in exactly the same place (same x-index) the element of the configuration matrix with one.

2. Modify the y-values with the same x-index over all curves of the set of curves.

 If in the first step the element is occupied by one, then subtract one from all y values. If the element was not occupied in the first step, then do nothing. 3. Continue through steps 1 and 2 until all elements of the configuration matrix have been addressed.

The configuration procedure is illustrated stepwise in Figure 9c. Finally, I would like to comment on a few peculiarities of the Houghcore configurations. The configurations for straight lines always represent only straight sections depending on the width of the Houghcore. Longer straight line sections in the binary edge image may need to be composed of several delektierten straight sections. The resolution of the angle (or slope) of the straight line sections depends on the height of the Houghcore.

The circle configurations always represent circular arcs around the vertex of the semicircle. Only the largest y-index of the family of curves (smallest radius) represents a complete semicircle. The developed configurations can be used for the new Houghcore.

Revision of the Houghcore

A major disadvantage of Holland-Neil's FPGA implementation is the rigid configuration of the Houghcores. The delavlines must be parameterized before the synthesis and are then stored permanently in the hardware structures (Holland-Neil, p. 48-49). Changes during runtime (eg Houghcore size) are no longer possible. The new procedure should become more flexible at this point. The new Houghcore should be too completely reconfigured during runtime in the FPGA. This has several advantages. On the one hand, not two Houghcores (Type 1 and Type 2) need to be stored in parallel, and on the other hand, different configurations for straight lines and semicircles can be used. In addition, the Houghcore size can be changed flexibly during runtime.

Previous Hou gh c ore structure consist of a delay and a bypass and it is determined before the FPGA synthesis which path is to be used. In the following, this structure is extended by one multiplexer, another register for configuring the delay element (switching of the multiplexer) and a pipeline delay. The configuration registers can be modified during runtime. In this way, different configuration matrices can be imported into the Hougheore. By using the pipeline selectays, the synthesis tool in the FPGA has more freedom in implementing the Houghcore design, and higher clock rates can be achieved. Pipe-linedelays break through time-critical paths within FPGA structures. In Fig. 9d, the new design of the delay elements is illustrated.

Compared to the previous implementation after Katzmann and Holland-Neil, the delay elements of the new Houghcore have a somewhat more complex structure. An additional register is required for flexible configuration of the delay element and the multiplexer occupies additional logic resources (must be implemented in the FPGA in a LUT). The pipeline delay is optional. In addition to the revision of the delay elements, modifications were also made to the design of the Houghcore. The new Hougheore is illustrated in Fig. 9e,

In contrast to the previous Hougheore, a new notation should first be introduced. Due to the 90 ° rotated design in Fig. 9e, the signals of the original histogram, originally called "row sums", are now called "column sums". Each column of the Houghcore thus represents a curve of the group of curves. The new Hougheore can also be loaded with new configuration matrices during runtime. The configuration matrices are in the FPGA internal BRAM. are stored and loaded by a configuration logic. This loads the configurations into the concatenated configuration registers as a column-wise bit string (see Fig. 9d). The reconfiguration of the Houghcore requires a certain amount of time and depends on the length of the columns (or the number of delay lines). Each column element requires one clock cycle and there is a latency of a few clock cycles through the BRAM and configuration logic. The overall latency for reconfiguration is disadvantageous, but can be accepted for video-based image processing, usually If the video streams recorded with a CMOS sensor have horizontal and vertical blanking, the reconfiguration can be done easily in the horizontal blanking time. The size of the Houghcore structure implemented in the FPGA also dictates the maximum size possible for Houghcore configurations. When small configurations are used, they are vertically centered and aligned horizontally at column 1 of the Houghcore structure (see Figure 91). Unused elements of the Houghcore structure are all filled with delays. The correct alignment of smaller configurations is important for the correction of the x-coordinates (see Formulas (B l 7) to (B 1 9)).

The Houghcore is fed as before with a binary edge image that passes through the configured delay lines. With each processing step, the column sums are calculated over the entire Houghcore and compared with the sum signal of the previous column, respectively. If a column returns a higher total, the sum value of the original column is overwritten. As an output, the new Houghcore returns a column sum value and its associated column number. On the basis of these values, a statement can later be made about which structure was found (represented by the column number) and with which probability of occurrence it was detected (represented by the sum value). The output signal of the Houghcore can also be referred to as Houghraum or Akkumulatorraum. In contrast to the conventional Hough transformation, the Hough space of the parallel Hough transformation is present in the image coordinate system. This means that a sum value with associated column number is output for each bi-coordinate. To completely transform the output image, a Floughcore of type 1 and type 2 must each be run through from the non-rotated and rotated image. Thus, after the transformation not only column sum with associated column number, but also the Houghcore type and the orientation of the output image (not rotated or rotated) are present. In addition, different Houghcore sizes and configurations can be used for straight and semi-circles. Thus, in addition to the results mentioned, the curve type and the Houghcore size can be specified. In summary, a result data set of the new Houghcore is shown in the following table. In the case of the parallel Hough transformation, such a data record is produced for each pixel of the output image.

Besehreibung

x-coordinate! Delayed according to the length of the Houghcore structure. An ingenious correction of the x-coordinate can be done i y-coordinate It is given according to the height of the _houghcore-structure

 Outline of the result data set created for each pixel of the output image in the parallel Hough transformation with a revised Houghcore structure. In contrast to. binary and threshold-based output of the Houghcore from Katzmann and Holland-Neil, the new Houghcore structure generates significantly more output data. Since such a data amount is difficult to handle, a method of feature extraction is presented, which significantly reduces the result data set. Type 2 Houghcore and picture rotation

In the parallel Hough transformation, the need for image rotation and the peculiarities of the type 2 Houghcore have already been discussed. In the parallel Hough transformation, the Ausgaiigsbild has to go through the Houghcore four times. This is necessary so that straight lines and semicircles can be detected in different angular positions. If you only work with a type 1 Houghcore, you have to process the image in the starting position and rotated by 90 °, 1 80 ° and 270 °. The addition of the Type 2 Houghcore eliminates 180 ° and 270 ° rotation. If the non-rotated output image is processed with a type 2 Houghcore, this corresponds to a processing processing of the 180 ° rotated output image with a Type 1 Houghcorc. The situation is similar with the rotation around 270 °. This can be replaced by processing the 90 ° rotated image with a Type 2 Houghcore. For FPGA conversion, the omission of additional rotations has a positive effect, since image rotations can normally only be solved with the help of an external memory. Depending on the hardware used, only a certain bandwidth (maximum possible data rate) is available between FPGA and memory chip. When using a Type 2 Houghcore, the bandwidth of the external memory module is only occupied by a rotation of 90 °. In the past implementation of Holland-Neil, it was necessary to deposit a type 1 Houghcore and a type 2 floughcore in the FPGA. With the revised Houghcore design, it is now also possible to drop the Houghcore structure once in the FPGA and load Type I or Type 2 configurations. This new functionality makes it possible to fully transform the source image with just one houghcore and just one image rotation.

It remains to be considered that when processing with only one Houghcore four times the data rate in Houghcore accrues. For a video stream with öOfps and VGA resolution, the pixel data rate is 24Mhz. In this case, the Houghcore would have to be operated at 96Mhz, which is already a high clock speed for a Spartan 3 generation FPGA. In order to optimize the design, work should increasingly be done with pipeline telelays within the floughcorc structure.

Feature of sextraction

The feature extraction works on the records from the previous table. These data sets can be summarized in a feature vector (B! 6). The feature vector can also be referred to below as a Hough feature.

MV = iMV _x , MVY VQ, MV KS, MV _N> MV _c i, MV _A ]

 (B 16)

A feature vector consists of the x and y coordinates for the found feature (MV _X and MV _y ), the orientation MVo, the curvature MV _/ cs, the frequency MV _H , the houghcore size MVQ.I and the type of found Structure MVA. The detailed meaning and the value range of the individual elements of the Merkmaisvektors can be seen in the following table.

MVx and MVy \ Both coordinates are each jijs to the size of the output image

 Elements of the Houg feature vector, their meaning and value range »

 Calculation of the orientation depending on the image rotation and the Houghcore type used for the transformation.

It can be seen from the above tables that the two elements MV ₀ and M ^' VKS have different meanings for straight lines and semicircles. In the case of straight lines, the combination of orientation and curvature strength forms the position angle of the detected straight line section in the angle! from 0 ° to 180 °. The orientation addresses an angular range and the curvature strength stands for a concrete angle within this range. The larger the Houghcore (the more 11 o ugheo re- palte n are present), the more finer is the angular resolution. In semicircles, the orientation is the position angle or the orientation of the semicircle. Semicircles can be detected by principle only in four orientations. The curvature is in semicircular configurations for the radius.

In addition to the orientation MV ₀ and the curvature MVKS is at the coordinates

(MVx and MVy) another special feature to note (see Fig. 9g). In the case of straight lines, the coordinates should always represent the center point and, in the case of semicircles or bends, always the vertex. With this specification, the y-coordinate can be corrected according to the implemented Houghcore structure and is independent of the size of the configuration used for transformation (see Fig. 9f). Similar to a local filter, the y-coordinate is specified vertically centered. For the x-coordinate, a relationship is established through the Houghcore column that yielded the hit (in the feature vector, the Houghcore column is placed under the designation MV _KS ), depending on the Houghcore type and the image rotation, arbitrary calculation rules for three different Cases are specified. For a type I Houghcore, the non-rotated and the rotated output image are each referenced to formula (B 1 7). If a Houghcore of type 2 exists, use formula (B 18) or formula (B 1) depending on the image rotation.

MV _r = MV _X + floor

(B 1 7)

MV. = Image width MV + floor

(B 1 8)

MV. = BiMhmüe,.,., "" - j MV t

(B 1)

The statement floor rounds off the fractional rational number, in the FPGA this corresponds to the simple descending of the binary decimal numbers. After the orientation has been determined and the coordinates of the Hough feature have been corrected, the actual feature extraction can take place. For feature extraction, three thresholds are used in combination with a non-maximum suppression operator. The non-maximum-suppression operator differs in straight lines and semicircles. A minimum MV _{k: s} and maximum curvature strength MV _{K are} specified via the threshold values and a minimum frequency MV _{H ^ is} set. The non-maximum suppression operator can be considered a 3x3 local operator (see Figure 9h). A valid semicircle feature (or curvature) always arises when the condition of the nms operator in (B23) is satisfied and the thresholds are exceeded according to formulas (B20) to (B22).

(B20)

MV *,> MV _K.

 (B21)

MV '> MV,

 (B22)

MV "A MV" M "MV" MV "MV" MV "MV" MV,

 (B23)

Non-maximum suppression suppresses Hough features that do not represent local Maxim a in the frequency domain of the feature vectors. This suppresses Hough features that do not contribute to the sought-after structure and are irrelevant to post-processing. The feature extraction would only be parameterized via three thresholds that can be sensibly adjusted in advance. A detailed explanation of the threshold values can be found in the following table.

Schwllllwert Spelling j Comparable parameter of the method according to Katzrnann

MV, Threshold for a minimum frequency, ie a column Hough Threshing value, which must not be undershot.

MV _K threshold for an m in imale curvature of the Hough feature. ! Bottom Line

 In Houghcores, with straightforward, rati fi ed, the

I threshold on the angle area detected by the Hougheore. MV K, S, behaves like MV _KS only for a maximum curvature, top-line

Detailed description of the three Scrr ellwcTte for extracting Hough features from the flight compartment. Compared to the Katzmann method, parameters with a similar function are given. For straight lines, a non-maximum suppression operator of size 3x3 (see Fig. 9h) can also be derived. There are some special features to consider. Unlike the curvatures, the sought-after structures in the straight line sections do not arise on a continuous several axima along the binary Kantenveriaufes. The non-maximum suppression can thus be based on the method in the Canny edge detection algorithm. Depending on the Houghcore type and the detected angle range, three cases can be distinguished (see Fig. 91 in combination with the above table). The case distinction applies to both rotated and non-rotated output bii, since the inverse transformation of rotated coordinates occurs only after non-maximum suppression. Which nms operator to use depends on the Houghcore type as well as the angle range. The angular range that a floughcore provides with straight line configurations is divided by the angular range bisector. The angular range bisector can be given as a Houghcore-Spaite (decimal broken) (MV _{KSI IH} ). The mathematical relationship depending on the Houghcore size is described by formula (B24). The range of the flough-fcature is based on the Houghcore-Spaite that delivered the hit (MVKS), which can be directly compared to the angular range-bisecting Houghcore-Spaite.

(B24)

If one has opted for an operator, the condition can be queried via the respective nms-oprator, similar to the non-maximum suppression for curvatures (formulas (B25) to (B27)). If all conditions are fulfilled and if the threshold values according to the formulas (B20) to (B22) are additionally exceeded, the Hough feature can be adopted at position »w.5 ^' 2.2.

Houghcore type [angle range nms operator condition

Area la A MF _K <MV _K;

Area ί b 7 MV KS> MV. Range 2a 2 A 1 V _KS <

Area 2b 2 B 3 MV _KS >

 Decision for an nms operator depending on the Houghcorc type and the angle range in which. the hit has occurred.

 (B27)

The conclusion of the feature extraction is the backward rotation of the x and y coordinates of noted Hough features. For postprocessing, these should be present again in the image coordinate system. The return trim information is to be executed whenever the rotated output image is processed, regardless of the curve type (whether straight lines or curves). The mathematical relationship is described in formulas (B28) and (B29). Image width refers to the width of the non-rotated output image ,

(B28)

MV _X = image width - MV _ym

 (B29)

With feature extraction hats it is possible to reduce the result data of the parallel Hough transformation to a few points. These can then be transferred to the postprocessing as a feature vector.

The above described embodiments are merely illustrative of the principles of the present invention. It will be understood that modifications and variations of the arrangements and details described herein will be apparent to others of ordinary skill in the art. Therefore, it is intended that the invention be limited only by the scope of the appended claims and not by the specific details presented with the description and illustration of the embodiment herein.

Claims

 P at first sayings

A 2D image analyzer (700), comprising: an image scaler (702) adapted to receive an image (710) having a searched pattern (711) and the received image (710) in accordance with a scaling factor to scale: an image generator (704) configured to generate an overview image (712) having a plurality of copies of the received and scaled image (710'-710 "" "'"), each copy being incremented by one different scaling factor is scaled; and a pattern finder (706) configured to compare a predetermined pattern with the plurality of received or scaled images (710'-710 "" "" ') within the overview image (712) and information regarding a position in which a match between the searched pattern and the predetermined pattern is maximum, the position relating to a respective copy of the received and scaled image (710'-710 "" "" '),

The 2D image analyzer (700) of claim 1, wherein each scaled image (710) is associated with a respective position in the overview image (712) according to the respective scaling factor.

A 2D image analyzer (700) according to claim 1 or 2, wherein the respective position is calculable by an algorithm having a spacing between the scaled images (710 ^ 710 ") in the overview image (712), a spacing of the scaled ones Images (710'-710 "" "" ') to one or more of the boundaries of the overview image (712) and / or other predefined conditions.

The 2D image analyzer (700) of any one of claims 1 to 3, wherein the pattern fInder (706) comprises a feature transformer for transforming one or more features and a classifier configured to form patterns in the overview image (712 ) or patterns in the survey «- image (712), which is converted into a feature space by means of the feature transformator.

5. The 2D image analyzer (700) of claim 4, wherein the feature transformer is configured to extract features, the feature transform based on a Hough algorithm, census algorithm, gradient algorithm, algorithm-based algorithm, and / or orreiations-based algorithm is based.

6. 2D image analyzer (700) according to claim 4 or 5, wherein the classifier is adapted to the classification based on a Census-transformed version of the overview image (712) or on the basis of a translated into a feature space version of the overview image (712) or on the basis of a version of the overview image (712) transferred to a gradient image.

7. A 2D image analyzer (700) according to claim 6, wherein the pattern finder (706) is configured to generate one or more local maxima in the census version of the overlay image or in the version of the overview image translated into the feature space (712) or in the version of the overview image (712) transferred to the gradient image, wherein a position of a local maximum determines the position of the identified predetermined pattern in the respective copy of the received and scaled image (710'-710 "" "" ).

A 2D image analyzer (700) according to claim 6 or 7, wherein said pattern member (706) has a sum age and / or a maximum filter used to smooth the position of the identified predetermined pattern in the respective copies of the received and scaled image (706). 710'-710 "" "" ') and / or for smoothing and position-correcting the local maxima in the classified overview image (702).

A 2D image analyzer (700) according to any one of claims 4 to 8, wherein the pattern manager (706) performs the feature transformation and classification for each received and scaled image (710'-710 """"') to determine the location of the image local maximum indicative of the position of the identified predetermined pattern in the respective copy of the received and scaled image (710'-710 """"') based on averaging.

10. The 2D image analyzer (700) according to one of the claims 4 to 9, wherein the position of the identified predetermined pattern relates to a reference point of a face, an eye or in general to the searched object,

1 1. The 2D image analyzer (700) of any one of claims 1 to 10, wherein the 2D image analyzer (700) is embedded in an embedded processor, programmable

Logic element or completely custom chips is implemented.

12. A 2D image analyzer (700) according to any one of claims 1 to 11, wherein the 21 ⁾ image analyzer (700) is connected to a processing unit having a selective adaptive data processor configured to generate a plurality of sets of Receive values, each set is assigned to a respective sample, and comprising a filter processor, wherein the filter processor is formed. to output plausible sentences based on the received sentences, and so that an implausible sentence is replaced by a plausible sentence.

13. The 2D image analyzer (700) of claim 1, wherein the 2D image analyzer (700) is connected to a processing unit having a D-image analyzer (700) configured to form at least a first set of Image data determined based on a first image (10) and another set of image information determined based on the first image (710) or another image (710), the first image (710) depicting a pattern of a three-dimensional object from a first perspective into a first image plane, and wherein the further set comprises relative information describing a relationship between a point of the three-dimensional object and the first image plane, the 3D image analyzer comprising; a position calculation means arranged to determine a position of the pattern in a three-dimensional space based on the first sentence, a further sentence determined on the basis of the further image (71 0), and a geometric relationship between the perspectives of the first and second views of the further image (710) or the position of the pattern in the three-dimensional space based on the first sentence and a statistically determined relationship. to calculate between at least two characteristic features in the first image (71 0) to each other; and an orientation calculator configured to calculate two possible 3-D view vectors, and from the two possible 3D view vectors

To determine a 3D BHck vector according to which the pattern is aligned in the three-dimensional space, wherein the calculation and determination based on the first sentence, the further sentence and on the calculated position of the pattern.

A 2D image analysis system comprising the 2D image analyzer (700) of claim 10, wherein the 2D image analyzer (700) is coupled to an image analysis system for capturing and / or tracking a pupil having a first Hough path for a first camera and a processing unit for reprocessing the first Hough path results, the first Hough path comprising a Hough processor having: a pre-processor (102) configured to generate a plurality of samples (1 12a , 1 12b, 1 12c), each comprising an image (710), and to rotate and / or mirror the image (710) of the respective sample (1 12a, 1 12b, 1 12c) and a plurality versions (1 12a \ 1 12a ", 1 12a '", 1 12a "") of the image (710) of the respective sample (U2a, 1 12h, 1 12c) for each sample (1 12a, 1 12b, 1 1.2c ) issue; and Hough transform means (104) arranged to generate a predetermined searched pattern in the plurality of samples (1 12a, 1 12b, 1 12c) based on the plurality of versions (1 12a ', 1 12a ", 1 12a "', 1 12a" "), whereby a characteristic of the Hough transformation device (104) which is dependent on the searched pattern can be adapted. wherein the processing unit comprises means for analyzing the detected patterns and outputting a set of geometry parameters describing a position and / or geometry of the pattern for each sample,

A 2D image analysis system comprising the 2D image analyzer (700) of claim 10 and an evaluation device, the evaluation device configured to detect a lack of response of the eye. A method of analyzing a 2D bi (710), comprising the steps of;

Scaling a received image (710) having a searched pattern according to a scaling factor;

Generating an overview image (712) having a plurality of copies of the received and scaled image (710'-710 "" "" '), each copy scaled by a different scaling factor: and

Comparing a predetermined pattern with the plurality of patterns of the plurality of received and scaled images (710'-710 """"') within the overview image (712), and outputting information regarding a position at which a match between the searched Pattern and the preestablished pattern, where the position refers to a respective copy of the received and scaled image (710'-710 ⁴ """").

A computer-readable digital storage medium having stored therein a computer program with program code for execution as it runs on a computer, an embedded processor, a programmable logic device or a custom chip, the method of claim 16.