DE102017201603A1 - Method and control unit for determining the position of a point of view - Google Patents

Abstract

A method (300) for determining the position of a viewpoint (122) based on image data (112) comprising an image (203) of a head (120) is described. The method (300) comprises determining (301) a plurality of blocks (201) of the image (203), wherein a block (201) comprises a subregion of the image (203). In addition, the method (300) comprises mapping (302) the plurality of blocks (201) to a corresponding plurality of feature values (211) of K different feature maps (210) using K different convolution weight vectors β (202), k = 1, ..., K and with K> 1. The K different convolution weight vectors β (202) were determined by means of a machine-learning method on the basis of training data. In addition, the method includes determining (303) the position of at least one viewpoint (122) based on the K distinct feature maps (210).

Description

The invention relates to a method and a corresponding control unit for detecting aspects and for determining the position of aspects, in particular in order to determine the condition of a driver and / or the head posture of a driver of a vehicle.

In a vehicle, functions may be provided depending on the condition of a driver of the vehicle and / or depending on the head position or gaze direction of the driver of the vehicle. For example, an at least partially autonomous driving function can be activated automatically when it is detected that the driver of a vehicle has averted his eyes from traffic.

The present document deals with the technical problem of providing a method and a control unit, by means of which the aspects of a person can be identified and located in an efficient, reliable and robust manner, in particular in order to determine the condition and / or the head posture of a driver of a vehicle ,

The object is solved by the features of the independent claims. Advantageous embodiments are described i.a. in the dependent claims. It should be noted that additional features of a claim dependent on an independent claim without the features of the independent claim or only in combination with a subset of the features of the independent claim may form an independent invention independent of the combination of all features of the independent claim, the subject of an independent claim, a divisional application or a subsequent application. This applies equally to technical teachings described in the specification, which may form an independent invention of the features of the independent claims.

According to one aspect, a method for determining the position of at least one viewpoint based on image data is described, the image data comprising an image of a head. Typically, the location is determined from several viewpoints (e.g., 20, 30, or more points of view). The position can indicate the (Cartesian) coordinates of a point of view within the image. On the basis of the position from several points of view, the head posture of a person, in particular a driver of a vehicle, can be determined.

The method includes determining a plurality of blocks of the image, wherein a block comprises a portion of the image. An image may e.g. 100 × 100 or more pixels. On the other hand, a block may be e.g. 3 × 3 or 5x5 pixels. The blocks may partially overlap.

In addition, the method includes mapping the plurality of blocks to a corresponding plurality of feature values of K different feature maps using K different convolution weight vectors β _{conv k} , k = 1, ..., K, and K> 1 thus K different convolutional weight vectors β _{conv k} are provided to analyze an image in different ways and thus to provide K different feature maps describing different aspects of the image.

In particular, the K different convolution weight _vectors β _{conv k} can be designed to assign a block K to different image structure types or to determine whether a block has one of K different image structure types. The different image structure types may include, for example, one or more edges and one or more curves (eg with different orientations in each case). The different feature maps (for the different image texture types) may then display the positions of different image structures within an image.

The K different convolution weight _vectors β _{conv k} may have been determined by means of a machine learning method based on training data. In particular, the K convolution weight vectors β _{conv k may be} weight vectors of K artificial neural networks, the image being an input layer and the K feature maps representing the output layers of the K neural networks. In addition, the artificial neural networks may have one or more hidden intermediate layers. In particular, the convolution weight _vectors β _{conv k} may include output weights of an extreme learning machine, ELM. Furthermore, the machine-learning method for determining the convolution weight _vectors β _{conv k may} include an Extreme Learning Machine, ELM, Auto Encoder, AE.

The extreme learning machine, ELM, is used eg in Z. Bai, G.-Bin. Huang, LLC Kasun and CM Vong, "Local Receptive Fields Based Extreme Learning Machine," IEEE Computational Intelligence Magazine, vol, 10, no. 2, 2015 described. The Extreme Learning Machine, ELM, Auto Encoder, AE, is used eg in LLC Kasun, H. Zhou, G.-B. Huang, and CM Vong, "Representational Learning with Extreme Learning Machine for Big Data, "IEEE Intelligent Systems, vol. 28, no. 6, pp. 31-34, 2013 described. These documents are incorporated herein by reference.

In addition, the method includes determining the position of at least one viewpoint based on the K different feature maps. By using different convolution weight _vectors β _{conv k} to analyze different portions of an image, the position of a viewpoint can be determined in an efficient and precise manner.

The convolution weight vectors can be determined in particular by the following formula: $${\beta }_{cOnv}={\left(\frac{I}{C}+H^{T}H\right)}^{-1}{H}^{T}T$$

Where [β β _{conv 1,} ..., _{conv K]} is the matrix β _conv = the summary of K different convolutional weight vectors β _{k conv;} I is an L × L unit matrix; C is a regularization term; and T is a matrix of training blocks x _n , where n = 1, ..., N, a plurality of training images. Thus, N training blocks can be provided to train the convolution weight vectors.

Further, H = [h (x ₁ ), ..., h (x _N )] ^T , where h (x _n ) L are initial values of L neurons of a hidden layer of neurons for the training gan x _n . L can be eg the number of pixels in a training block x _n .

sein, wobei h_l(x_n) = g(a_l · x_n + b_l), mit l = 1, ...,L, eine Aktivierungsfunktion des l^ten Neurons der versteckten Schicht ist; und wobei a_l zufällige Gewichtungen und b_l ein zufälliger Schwellenwert sind. So wird somit ein effizientes und präzises Anlernen der Faltungs-Gewichtsvektoren ermöglicht. Die Faltungs-Gewichtsvektoren können im Vorfeld zur Ermittlung der Position zumindest eines Gesichtspunktes anhand der o.g. Formeln ermittelt worden sein.In addition, can $$H\left(x_n\right)={\left[H_1\left(x_n\right)\text{.}...\text{.}{H}_L\left(x_n\right)\right]}^{\text{T}}$$

, where h _{l (n} · x + b _l a _l) is to be (x _n) = g, with l = 1, ..., L, activation function of the l ^th neuron of the hidden layer; and where a _{l are} random weights and b _{l is} a random threshold. Thus, an efficient and precise learning of the convolution weight vectors is thus made possible. The convolution weight vectors may have been determined in advance to determine the position of at least one aspect using the above-mentioned formulas.

wobei c_k,i(x) der i^te Merkmalswert der k^ten Merkmals-Karte ist; wobei x_l ein l^ter Bildpunkt des Blockes x ist; und wobei β_{conv k,l} ein l^tes Faltungs-Gewicht des Faltungs-Gewichtsvektors β_{conv k} ist.The mapping of a block x by means of a convolution weight _vector β _{conv k} may comprise $$c_{k\text{.}i}\left(x\right)={\displaystyle \underset{l=1}{\overset{L}{\Sigma }}{x}_l{\beta }_{cOnvk\text{.}l}}$$

where c _{k, i} (x) is the i ^th feature value of the k ^th feature map; where x _{l is} an ^lth pixel of the block x; and wherein β _{conv k, l} is an l ^th convolution weight of the convolutional weight vector β _{conv k.}

Determining the position of at least one viewpoint may include, for each of the K distinct feature maps, determining a corresponding summary map, each having a plurality of summary values. In this case, a summary value of the k ^th summary map can be determined on the basis of a standard, in particular an L2 standard, of a feature block of feature values (eg with 2x2 feature values) of the k ^th feature map. Thus, the method can be caused to be translation invariant.

Determining the position of at least one aspect may comprise, for each of the K different feature maps, determining a connection vector having a reduced number of connection values compared to the number of feature values and the number of summary values, respectively. In this case, the connection values of a k ^th connection vector based on connection weight vectors P _{full k} , with k = 1, ..., K. The K different connection weight vectors β _{full k} may have been determined by means of a machine-learning method, in particular by means of an ELM AE, on the basis of the training data. Thus, relationships between the different image structures can be determined.

In addition, determining the position of at least one aspect may include determining a position vector based on the K connection vectors. In this case, the K connection _{vectors can be converted} into the position vector by a common output weight vector β _output . The output weight vector β _output may have been determined by means of a machine-learning method on the basis of the training data and on the basis of labels for the training data, the labels indicating actual positions of viewpoints in the training data. By means of the output weight vector β _output , the image structures of the individual feature maps or connection vectors can be combined in order to robustly detect aspects and determine their position.

According to a further aspect, a control unit for determining the position of a viewpoint on the basis of image data is described, the image data comprising an image of a head. The control unit is set up to determine a plurality of blocks of the image, one block comprising a partial area of the image. In addition, the control unit is set up by means of K different convolution weight _vectors β _{conv k} the plurality of Mapping blocks to a corresponding plurality of feature values of K different feature maps, where k = 1, ..., K and K> 1 (where K is an integer). The K different convolution weight _vectors β _{conv k were determined} by means of a machine-learning method on the basis of training data. The control unit is further configured to determine the position of at least one viewpoint based on the K different feature maps.

According to a further aspect, a vehicle (in particular a road motor vehicle, for example a passenger car, a truck or a motorcycle) is described, which comprises the control unit described in this document (for example, to determine the head posture of a driver of the vehicle).

In another aspect, a software (SW) program is described. The SW program can be set up to run on a processor and thereby perform the method described in this document.

In another aspect, a storage medium is described. The storage medium may include a SW program that is set up to run on a processor and thereby perform the method described in this document.

It should be understood that the methods, devices and systems described herein may be used alone as well as in combination with other methods, devices and systems described in this document. Furthermore, any aspects of the methods, devices, and systems described herein may be combined in a variety of ways. In particular, the features of the claims can be combined in a variety of ways.

• Figur 1a beispielhafte Komponenten eines Fahrzeugs;
• Figur 1b beispielhafte Gesichtspunkte eines Gesichts eines Fahrers eines Fahrzeugs;
• 2 eine beispielhafte Architektur eines Algorithmus zur Ermittlung der Position von Gesichtspunkten; und
• 3 ein Ablaufdiagramm eines beispielhaften Verfahrens zur Ermittlung der Position eines Gesichtspunktes.

Furthermore, the invention will be described in more detail with reference to exemplary embodiments. Show

• Figure 1a exemplary components of a vehicle;
• FIG. 1b shows exemplary aspects of a face of a driver of a vehicle;
• 2 an exemplary architecture of an algorithm for determining the position of viewpoints; and
• 3 a flowchart of an exemplary method for determining the position of a viewpoint.

As set forth above, the present document is concerned with the efficient and reliable determination of the position of one or more aspects of a human, in particular a driver of a vehicle, based on image data. In this context, Fig. La shows exemplary components of a vehicle 100 , The vehicle 100 comprises an image sensor 101, eg an image camera, which is set up, image data 112 concerning the head of an occupant, in particular the driver, of the vehicle 100 capture.

The image data 112 can take a picture of the head 120 of the occupant of the vehicle 100 (see 1b ).

The vehicle 100 includes a control unit 101 that is set up the image data 112 analyze. In particular, the control unit 101 be set up based on the image data 112 the position of one or more points of view 122 on the head 120 , in particular on the face, of the occupant. The points of view 122 can be part of a distinctive area 121 , eg part of an eye, a nose, a mouth, an eyebrow, etc., of the face 120 be. In particular, the positions may have several aspects 122 be determined (such as in 1b shown). So can a model of the head 120 be created by the occupant, for example, the condition of the occupant and / or the attitude of the head 120 to investigate. The control unit 101 can then, for example, depending on the attitude of the head 120 , one or more actuators 103 of the vehicle 100 to provide a driver assistance system.

To determine the position of a point of view 122 Various image analysis algorithms can be used. This document describes an algorithm that uses trained artificial neural networks to capture the image data 112 concerning the head 120 assign a human to a seed vector that is the position of one or more points of view 122 displays. 2 illustrates the algorithm described in this document 200 ,

The image data 112 represent a picture 203 with a plurality of pixels 204 , in particular with a matrix of pixels 204 through which the head 120 of a human being. The picture 203 can be divided into a plurality of blocks x _j 201. The blocks 201 have, for example, a size of r × r, where r = 10, 5 or 3. The blocks 201 may at least partially overlap. For example, the blocks 201 may be formed by starting from one corner of the image 203 (eg top left), by shifting to the right by one pixel at a time 204 and / or by moving down each a pixel 204 a new block 201 is generated. This results in a picture 203 with the dimensions d × d, (d-r + 1) × (d-r + 1) blocks 201 ,

wobei x_i+m-1,j+n-1 die Pixel 204 eines Blockes 201 und β_{conv k m,n} die Faltungs-Gewichte eines k^ten Faltungs-Gewichtsvektors sind. In der o.g. Formel werden die Zeilen i und die Spalten j eines Blocks betrachtet. Alternativ können die Blöcke 201 als eindimensionale Vektoren dargestellt werden (durch Aneinanderreihung der einzelnen Spalten).A block 201 may converte 202 by a convolution weight vector β _{conv k} 202 on a feature map 210 be imaged. This includes a feature card 210 for each of the multitude of blocks 201 a feature value c _{i, j} (x) 211. It can thus be a feature map 210 with a plurality of feature values c _{i, j} (x) 211. The characteristic values 211 from the corresponding blocks 201 are determined by convolution or convolution with the filter formed by the convolution weight _vector β _{conv k} 202, as $$\begin{array}{c}{c}_{i\text{.}j}\left(x\right)={\displaystyle \underset{m=1}{\overset{r}{\Sigma }}{\displaystyle \underset{n=1}{\overset{r}{\Sigma }}{x}_{i+m-\mathrm{1,}j+n-1}\cdot {\beta }_{cOn{v}_{k_{m\text{.}n}}}}}\\ i\text{.}j=\mathrm{1,}...\text{.}\left(d-r+1\right)\end{array}$$

where x _{i + m-1, j + n-1 are} the pixels 204 a block 201 and β _{conv k m, n} are the convolution weights of a k ^th convolution weight vector. In the above formula, the lines i and the columns j of a block are considered. Alternatively, the blocks 201 represented as one-dimensional vectors (by juxtaposing the individual columns).

There can be K different convolution weight vectors 202 , β _{conv 1} to β _{conv K} used to get out of the large number of blocks 201 of a picture 203 K different feature cards 210 to investigate.

By identifying different feature cards 210 different distinctive structures, such as edges or curves, can be detected. Through the feature works 211 The different feature maps 210 may thus have different structures within the original image 203 are displayed. For example, by a first feature card 210 Structures of a first structure type (eg, edges in a first direction) and a second feature map 210 Structures of a second structure type (eg, edges in a second direction) are displayed.

The application of different folding weight vectors 202 on subareas, ie on blocks 201 , a picture 203 makes it possible to efficiently consider the fact that between adjacent pixels 204 of an image 203 typically a relatively high correlation and between more distant pixels 204 a picture 203 typically has a relatively low correlation.

The convolution weight vectors 202 can be selected at random, eg by means of a certain probability distribution. On the other hand, the quality of the recognition of viewpoints 122 be substantially increased when the convolution weight vectors 202 be learned on the basis of training data. This can be advantageously used the so-called Extreme Learning Machine (ELM) Auto Encoder (AE), for example, in LLC Kasun, H. Zhou, G.-B. Huang, and CM Vong, "Representational Learning with Extreme Learning Machine for Big Data," IEEE Intelligent Systems, vol. 28, no. 6, pp. 31-34, 2013 , is described. In particular, the equation (6) written in the above-mentioned document can be used to construct convolution weight vectors 202 based on training data to determine.

wobei h_p,q,k in der obigen Gleichung der Zusammenfassungswert 221 p, q für die k^te Zusammenfassungs-Karte 220 ist. Es ergeben sich somit K Zusammenfassungs-Karten 220 mit Zusammenfassungswerten 221. Durch die Ermittlung von Zusammenfassungs-Karten 220 kann der Algorithmus 200 skalen- und translations-invariant gemacht werden.The in 2 The algorithm illustrated comprises a summarizing step in which a plurality of feature values 211 a feature card 210 to a summary value 221 be summarized. In particular, feature blocks 214 from e × e characteristic values 211 are formed (eg e = 3 or 2), where from the feature values 211 a feature block 214 each a summary value 221 is determined. The feature blocks 214 can overlap at least partially. A summary value 221 may be obtained from the feature values by means of a norm, eg an L1 or L2 norm 211 a feature block 214 be determined, eg as $$H_{p\text{.}q\text{.}k}=\sqrt{{\displaystyle \underset{i=p-e\text{/}2}{\overset{p+e\text{/}2}{\Sigma }}{\displaystyle \underset{j=q-e\text{/}2}{\overset{q+e\text{/}2}{\Sigma }}{c}_{i\text{.}j\text{.}k}^2}}}$$

where h _{p, q, k} in the above equation is the summary value 221 p, q for the k ^th summary card 220 is. This results in K summary maps 220 with summary values 221 , By identifying summary cards 220 can the algorithm 200 scales and translations are made invariant.

In a connection step, values 221 . 211 the individual resume cards 220 or feature cards 210 be connected to each other. For this purpose, for each card 220 . 210 Compound weight _vectors P _fullk 222 can be detected. The connection weight _vectors P _fullk can advantageously by means of the _{above-mentioned} ELM-AE method based on maps 220 . 210 be determined from training data with a variety of images 203 were determined. The connection weight vectors 222 may be the values of a map 220 . 210 to a reduced number of connection values 231 a connection vector 230 depict. This results in K connection vectors 230 for the K cards 220 . 210 ,

Finally, the K connection vectors 230 be converted by the output weights of an output weight vector β _output 232 into a position vector 240 indicative of the positions of one or more viewpoints 122. The output weights can be determined by means of linear regression or by means of the above-mentioned ELM-AE method based on training data and using labels showing the actual positions of one or more aspects 122 in the training data show, be learned.

By the in 2 The algorithm depicted may efficiently and precisely determine the position (eg, the x / y coordinates) of one or more aspects 122 of a human on the basis of image data 112 be determined. Thus, the head posture of a person can dependably based on the position of viewpoints 122 be determined (even when changing lighting situations)

3 shows a flowchart of an exemplary method 300 for determining the position (in particular the coordinates) of a viewpoint 122 on the basis of image data 112 , where the image data 112 a picture 203 a head 120 of a human being. In particular, the image data 112 the picture 203 a head 120 a driver of a vehicle 100 include. Based on the position of one or more points of view 122 Of the head 120 the driver may be on an attitude of the head 120 , For example, in a line of sight, the driver to be closed. This information can be used to control a driver assistance system. The procedure 300 is a computer-implemented method, for example, on a control unit 101 a vehicle 100 can be executed.

The procedure 300 includes determining 301 a variety of blocks 201 of the picture 203 , For example, with an offset of one pixel each 204 in the horizontal direction and / or in the vertical direction different blocks 201 from a picture 203 be formed. A block 201 thus includes a portion of the image 203 (eg a subarea with r × r pixels 204 ).

In addition, the process includes 300 the picture 302 of the plurality of blocks 201 to a corresponding plurality of feature values 211 from K different feature cards 210 using K different convolutional weight vectors β _{conv k} 202, k = 1, ..., K and K> 1. In particular, each block 201 the variety of blocks 201 with K different convolutional weight vectors β _{conv k} 202 to filter the feature values 211 of K different feature maps 210 to investigate. The K different convolution weight _vectors β _{conv k} 202 can be designed here, K different image structure types from a block 201 filter out. By using K different convolution weight vectors β _{conv k} 202, K can thus use different feature maps 210 for a picture 203 are each provided the position of different image structure types in the image 203 Show.

The K different convolution weight _vectors β _{conv k} 202 may have been determined by means of a machine-learning method on the basis of training data. In particular, an ELM-AE method may have been used to determine the convolution weight _vectors β _{conv k} 202. So, in an efficient and reliable way can analyze an image 203 in terms of different image structure types.

In addition, the process includes 300 the determining 303 the position of at least one point of view 122 based on the K different feature maps 210. For this purpose, an artificial neural network can be used that is set up, the K different feature maps 210 to a position vector 240 depicting the position, in particular the coordinates, of at least one point of view 122 displays. The artificial neural network can be trained on the basis of the training data and on the basis of labels for the training data. The labels show the actual position of the at least one aspect 122 for the training data.

It thus becomes an efficient and reliable determination of the position of viewpoints 122 allows.

The present invention is not limited to the embodiments shown. In particular, it should be noted that the description and figures are intended to illustrate only the principle of the proposed methods, apparatus and systems.

QUOTES INCLUDE IN THE DESCRIPTION

This list of the documents listed by the applicant has been generated automatically and is included solely for the better information of the reader. The list is not part of the German patent or utility model application. The DPMA assumes no liability for any errors or omissions.

Cited non-patent literature

• Z. Bai, G.-Bin. Huang, L.L.C. Kasun and C.M. Vong, "Local Receptive Fields Based Extreme Learning Machine", IEEE Computational Intelligence Magazine, vol, 10, no. 2, 2015. [0010]
• L.L.C. Kasun, H. Zhou, G.-B. Huang, and C.M. Vong, "Representational Learning with Extreme Learning Machine for Big Data," IEEE Intelligent Systems, vol. 28, no. 6, pp. 31-34, 2013 [0010, 0035]

Claims (10)

A method (300) for determining the position of a viewpoint (122) based on image data (112) comprising an image (203) of a header (120), the method comprising (300), - determining (301) a plurality of Blocks (201) of the image (203); wherein a block (201) comprises a portion of the image (203); Mapping (302) the plurality of blocks (201) to a corresponding plurality of feature values (211) of K different feature maps (210) using K different convolution weight vectors β _{conv k} (202), where k = 1, .. ., K and with K>1; wherein the K different convolution weight _vectors β _{conv k} (202) were determined by means of a machine learning method based on training data; and - determining (303) the position of at least one viewpoint (122) based on the K distinct feature maps (210). Method (300) according to Claim 1 where - the machine-learning method includes an Extreme Learning Machine, ELM, Auto Encoder, AE; and / or - a convolution weight vector β _{conv k} (202) comprises output weights of an extreme learning machine, ELM. Method (300) according to one of the preceding claims, wherein $${\beta }_{cOnv}={\left(\frac{I}{C}+H^{T}H\right)}^{-1}{H}^{T}T$$

where - β _conv = [β _{conv 1} , ..., β _{conv K} ]; I is an L × L unit matrix; - C is a regularization term; - T is a matrix with training blocks (201), x _n, with n = 1, ..., N, a plurality of training images (203); - H = [h (x ₁ ), ..., h (x _N )] ^T ; - h (x _n ) L are initial values of L neurons of a hidden layer of neurons for the training block (201) x _n ; and L is a number of pixels (204) in a training block (201) x _n . Method (300) according to Claim 3 , in which $$H\left(x_n\right)={\left[H_1\left(x_n\right)\text{.}...\text{.}{H}_L\left(x_n\right)\right]}^{\text{T}}$$

where - h _l (x _n ) = g (a _l * x _n + b _l ), where l = 1, ..., L, is an activation function of the ^lth neuron of the hidden layer; and - a _{l l} b are random weights and a random threshold. Method (300) according to one of the preceding claims, wherein - the K different convolution weight _vectors β _{conv k} (202) are designed to assign a block (201) to K different image structure types; and the image structure types in particular comprise: one or more edges and one or more curves. The method (300) of one of the preceding claims, wherein mapping (302) a block x (201) by means of a convolution weight _vector β comprises _{conv k} (202), $$c_{k\text{.}i}\left(x\right)={\displaystyle \underset{l=1}{\overset{L}{\Sigma }}{x}_l{\beta }_{cOnvk\text{.}l}}$$

where - c _{k, i} (x) is the i ^th feature value (211) of the k ^th feature map (210); - x _{l is} a l ^th pixel (204) of the block x (201); and - ß _{conv k, l} a l ^th convolution weight of the convolutional weight vector β _{conv k} (202). A method (300) according to any one of the preceding claims, wherein - determining (303) the position of at least one viewpoint (122) for each of the K distinct feature maps (210) comprises determining a respective summary map (220) each having a plurality of summary values (221); and - a merge value (221) of the ^short merge map (220) is determined based on a norm, in particular an L2 norm, of a feature block (214) of feature values (211) of the ^kth feature map (210) , The method (300) according to one of the preceding claims, wherein - determining (303) the position of at least one viewpoint (122) for each of the distinct K feature maps (210) comprises determining a connection vector (230) against a number of feature values (211) comprises a reduced number of connection values (231); - the connection _values (231) of a k ^th connection _vector (230) are determined using connection weight _vectors P _fullk (222), _where k = 1, ..., K; and - the K different connection weight _vectors P _fullk (222) were determined by means of a machine learning method based on training data. Method (300) according to Claim 8 wherein determining (303) the position of at least one viewpoint (122) comprises determining a Position vector (240) based on the K connection vectors (230); - K compound vectors (230) _output (232) is transferred into the position vector (240) by an output weight vector β; the output weight vector β _output (232) has been determined by means of a machine learning method on the basis of the training data and on the basis of labels for the training data; and - the labels indicate actual positions of viewpoints (122) in the training data. Control unit (101) for determining the position of a viewpoint (122) on the basis of image data (112) comprising an image (203) of a head (120), the control unit (101) being arranged, - a plurality of blocks (201 ) of the image (203); wherein a block (201) comprises a portion of the image (203); by means of K different convolution weight vectors β _{conv k} (202) to map the plurality of blocks (201) to a corresponding plurality of feature values (211) of K different feature maps (210), where k = 1, ..., K and with K>1; wherein the K different convolution weight _vectors β _{conv k} (202) were determined by means of a machine learning method based on training data; and determine the position of at least one viewpoint (122) based on the K different feature maps (210).

Images