Method and system for automatic steering of an agricultural vehicle

p0001The invention relates to a method for automatically steering an agricultural machine in a field along a surface to be machined, and to a corresponding steering system.

p0002In recent years, many features of agricultural vehicles which have previously been controlled manually by an operator, have been automated. To facilitate the work of the operator, different steering systems have been proposed, for example, direct the vehicle automatically across a field. You can facilitate the operator's work and enable it to focus on other important parts of his work, as a control of the utilization of working equipment of the vehicle. It could also be possible in the future to allow such vehicles drive without a driver over a field.

p0003It automatic steering systems have been described which are based on different principles. Non-contact systems can be divided into based on a distance measurement systems that attempt to detect the height difference between cut and uncut crop, and based on an imaging systems that are trying to differences in appearance between the crop on the cut and the uncut side the harvested crop or plant and the ground to detect.

p0004Examples based on a distance measurement systems can be found in the <patcit id="pcit0001" dnum="EP0887660A"><text>EP 0887660 A</text></patcit> and <patcit id="pcit0002" dnum="EP1271139A"><text>EP 1271139 A</text></patcit>,

p0005In the <patcit id="pcit0003" dnum="EP0887660A"><text>EP 0887660 A</text></patcit> is pivoted over an angle range addressed on the floor in front of a harvester laser distance measuring device. By means of their orientation and position of the corresponding to the rotational angle position of the point is determined at which the laser beam was reflected. In this way, a height profile of the soil is determined, which allows a swath or a crop material edge on the border of a harvested part of a cornfield and the as yet unharvested part to recognize and generate a steering signal.

p0006The <patcit id="pcit0004" dnum="EP1271139A"><text>EP 1271139 A</text></patcit> proposes in such a laser distance-measuring device prior to addition, based on the intensity of the reflected radiation to determine the amount of the male crop.

p0007A disadvantage of based on a distance measurement systems is that a relatively costly distance measuring device with a laser and electronics are required at run time recording.

p0008Examples based on an imaging systems that operate with one targeted at the box camera and electronic image processing, are in the <patcit id="pcit0005" dnum="DE3507570A"><text>DE 35 07 570 A</text></patcit>. <patcit id="pcit0006" dnum="EP0801885A"><text>EP 0801885 A</text></patcit>. <patcit id="pcit0007" dnum="WO9846065A"><text>WO 98/46065 A</text></patcit>. <patcit id="pcit0008" dnum="WO9617279A"><text>WO 96/17279 A</text></patcit>. <patcit id="pcit0009" dnum="JP01319878A"><text>JP 01319878 A</text></patcit>. <patcit id="pcit0010" dnum="JP03012713A"><text>JP 03012713 A</text></patcit>. <patcit id="pcit0011" dnum="US6278918B"><text>US 6278918 B</text></patcit>. <patcit id="pcit0012" dnum="US6285930B"><text>US 6285930 B</text></patcit>. <patcit id="pcit0013" dnum="US6385515B"><text>US 6385515 B</text></patcit>. <patcit id="pcit0014" dnum="US6490539B"><text>US 6490539 B</text></patcit> and <patcit id="pcit0015" dnum="US2002106108A"><text>US 2002/106108 A</text></patcit> to find.

p0009The <patcit id="pcit0016" dnum="DE3507570"><text>DE 35 07 570</text></patcit> discloses an automatic steering system for an agricultural vehicle to drive along one or a field cultivated in rows plants. The system derives the orientation of the rows of plants from an initial image. Then the image along lines is scanned, which extend parallel to the secondary orientation and the mean gray values ​​of each line are stored in a memory. While driving across the field, the field is scanned along the same lines. The results of the scanning of each line are compared with the values ​​in memory and the reference values ​​are maintained by appropriate steering signals.

p0010In the <patcit id="pcit0017" dnum="EP0801885A"><text>EP 0801885 A</text></patcit> is described with a camera, digitizes the image signal and an image processing is subjected to a harvester. The rows are each examined the position of a step function, at which the gray value changes gradually, towards. In each row, the pixel positions are determined with the highest probability of the step function. These positions are used to generate a steering signal.

p0011In the <patcit id="pcit0018" dnum="WO9846065A"><text>WO 98/46065 A</text></patcit> is also recognized on the basis of the information contained in the individual lines, whether they contain a Erntegutgrenze. Based on the information of all the lines is determined, in which line ends the field being harvested.

p0012The <patcit id="pcit0019" dnum="WO9617279A"><text>WO 96/17279 A</text></patcit> proposes a limit (z. B. Parallelogram) to put on an image taken by a camera image and measure inherent in objects with a brightness that is above a threshold to identify. The objects are subjected to a regression analysis, to determine their offset to a previously identified series, in turn, is used to generate a steering signal.

p0013The <patcit id="pcit0020" dnum="JP01319878A"><text>JP 01319878 A</text></patcit> refers to a steering system, are joined together at the surfaces of specific color in an image captured by a camera image to define a straight line. This line is regarded as a row of plants and is used for generating the steering signal.

p0014In the <patcit id="pcit0021" dnum="JP03012713A"><text>JP 03012713 A</text></patcit> is an automatic steering system for a tractor proposed in which a window of a recorded image is defined. This window displays a virtual line is defined on the arranged the plants. In this case, the image with a virtual row of plants is compared to generate a steering signal.

p0015The <patcit id="pcit0022" dnum="US6278918B"><text>US 6278918 B </text></patcit>proposes to divide the captured image into two or four sections. In the areas of a place and a direction of a row of plants is determined. The more visible row of plants is used to generate the steering signal. The decision whether a pixel representing a plant, or is not done, by so-called K-means clustering, in which a histogram of pixel values ​​is produced in which the x-axis shows the gray scale levels and the y-axis the number of pixels with the respective representing gray values. The pixels that represent likely the plants, therefore appear on the right side of the histogram. The algorithm divides the histogram generated from the camera image into a predetermined number K of classes whose intervals are each equally spaced. The data are processed recursively, to find the average gray value for the class. The boundaries between the classes are based on a predetermined metric or minimizing a cost function shifted recursively until shift the centers of classes by less than a given tolerance. Thereafter, a threshold for distinguishing between plants and the ground is set at the beginning of class K or class. The classification can be done through other clustering algorithms, such as sharing mechanism, self-organizing maps and neural logic or fuzzy logic. The<patcit id="pcit0023" dnum="US6285930B"><text>US 6285930 B</text></patcit>. <patcit id="pcit0024" dnum="US6385515B"><text>US 6385515 B</text></patcit> and <patcit id="pcit0025" dnum="US6490539B"><text>US 6490539</text></patcit> B also refer to this system.

p0016In the <patcit id="pcit0026" dnum="US2002106108A"><text>US 2002/106108 A</text></patcit> another system based on one camera and an image processing system steering system will be described. the inclination and the intersection of the edge of the standing rigging will be charged with the edge of the picture Based on the recorded image.

p0017In the described, based on an imaging systems a distinction is made on the basis of brightness and / or color of each pixel, whether they represent a plant or not. The plants are identified for some of these systems are connected by a virtual line, which serves to generate a steering signal. However, it is in some cases in which an automatic steering an agricultural machine is desirable, not simply or even nearly impossible to see from the color or brightness, whether a pixel belongs to a line to be followed or not. A typical example is a swath cut grass or lying on a harvested cornfield straw. The swath consists of the same material as standing on the field remaining stubble and therefore in many cases, more or less the same brightness and color. The imaging systems described are difficult to use here. Other systems described recognize plants in a field and are based on the series described by them to discharge, for example, chemicals. If no plants are present, these systems can not be used.

p0018In the introduction to the <patcit id="pcit0027" dnum="EP0801885A"><text>EP 0801885 A</text></patcit> (Page 3, first paragraph) mentions a possible method that uses a local two-dimensional Fourier transform operator as a base for a texture-based segmentation to determine a spatial frequency band with one major difference between the cut plants and stationary plants. It is stated that this method no clear indication of such a difference was found in initial tests. It has therefore been rejected.

p0019The problem of the invention underlying is seen to provide a method for automatically steering an agricultural machine and a corresponding steering system that is reliable and cost-effective manner.

p0020This problem is achieved by the teaching of claim 1, wherein in the further patent claims cite characteristics that further develop the solution in an advantageous manner.

p0021It is proposed a method and an apparatus for the automatic steering of an agricultural machine along a surface to be machined on a field. The machine is provided with a camera that is directed to the area in front of the machine, including the area to be processed. The camera provides at regular intervals two-dimensional images of the surface to be processed and the construction nearby areas of the field. a bitmap is generated from the image signal from the camera, which is further processed by an electronic image processing to generate a steering signal and to cause the machine automatically along the surface to be processed. An advantage of the use of an imaging system is over systems based on a distance measurement in the reduced cost.

p0022The invention proposes to derive a texture information with respect to the texture of the environment of each pixel from the evaluated pixel file. There is an information generated, which contains information about how the structure of the image in the neighborhood of the pixel looks, ie, whether it is, for example, regular or irregular, or the direction in which there existing details such. B. Ernteguthalme or clods, oriented are. Look the camera, for example on a mowed field with a windrow, the field shows mainly short vertical stalks and the swath longer straws with different orientations. the surface to be machined is analog in tillage of the adjacent field areas different because it more or less, for example, after plowing or cultivating large clods and a more uneven surface shows than before. This texture information is used as a basis for classification, whether each pixel of the surface to be machined to be assigned or not (binary decision). As not only the brightness of the pixel, but also detailed information of the classification are used as a basis as in the prior art, it is on a relatively secure basis. The position of the surface to be processed can thus be detected without problems. On the basis of the known position of the surface to be treated (or not and / or not yet to be processed surface) representing the pixel is formed a steering signal, which is supplied to the steering means of the machine.

p0023The texture information can be obtained in particular by a non-directional or directional texture analysis. In a non-directional texture analysis information to be considered in all directions with respect to the neighbors of the pixel. In a directional texture analysis only the neighbors are evaluated in certain directions. The direction-dependent texture analysis is particularly useful if preferably occurring in the image characteristics in certain directions. The directional or non-directional texture analysis can be performed for example by means of a gray value dependency analysis.

p0024Alternatively or additionally, the texture can be examined by means of a multispectral analysis. Here, a camera is used, which can detect at least two different wavelengths of light. The different wavelengths also the texture in the neighborhood of pixels can differ, so that one. Deciding whether the individual pixels each representing the surface to be machined or not, can lay additional information to reason As multispectral analysis, a color analysis, in particular, an RGB analysis are provided.

p0025The decision as to whether the individual pixels each representing the area to be processed or not is made by a so-called classifier. A suitable classifier is a neural network. The classifier is preferably subjected before starting work by means of a supplied thereto and a image information on parts of the image showing the surface to be machined, and areas, which show a non surface to be machined, a learning operation. The learning process can be carried out on a field in the production in the factory or before starting the work.

p0026The steering system according to the invention is particularly suitable for guiding the harvesting machine along provided as a swath the crop because it allows a good distinction between the stalks in the swath and on the field. Such harvesters are, for example, forage harvester with a pick-up as a crop intake, self-propelled or towed by a tractor balers and tractors with Trailer. The swath is detected and the harvester is guided along it. The steering system described can also be used on harvesters, which are moved along a crop material, such as combine harvesters, forage harvesters with corn harvesting devices or mowers, because the steering system is capable of, the long-standing stalks with ears on the top of the stubble on the harvested field to distinguish. Here, the harvesting machine is guided along the outer edge of the crop pick-up at the edge of the harvested crop.

p0027It should be noted that the image produced by the camera also contains information on the amount or rate of the male crop. This information can be derived from the contours of the crop and / or its color. The steering system can thus calculate the expected rate and create an automatic control of the vehicle speed based.

p0028Another possible use case of the steering system according to the invention is for example by means of a plow, a cultivator, a harrow, a roller or the like. Guided through tillage or seeding. As already mentioned above, the texture of the soil is different after processing of the previously existing texture. The steering system detects the boundary between the already machined part of the field and the unmachined part by the different textures and directs the agricultural machine at her at an optimal distance along, so that the outer limit of their working tool on the border between the processed part of the field and the unprocessed part is guided along.

p0029In the drawings, two embodiments of the invention described in more detail below are shown. It shows:<dl id="dl0001"><dt>Fig. 1</dt><dd>a forage harvester with an inventive automatic steering system in side view and in a schematic representation,</dd><dt>FIG. 2</dt><dd>a schematic side view of a tractor with an automatic steering system according to the invention and to which it has baler,</dd><dt>Fig. 3</dt><dd>a block diagram of the steering system,</dd><dt>Fig. 4</dt><dd>an example of an image taken by the camera,</dd><dt>Fig. 5</dt><dd>a flow chart according to which the processor operates the steering system,</dd><dt>Fig. 6</dt><dd>a schematic of a for the decision whether a pixel belongs to the field or to the swath, usable neural network,</dd><dt>Fig. 7</dt><dd>the operation of the neural network of <figref idrefs="f0006">figure 6</figref> in a learning phase,</dd><dt>Fig. 8</dt><dd>the operation of the neural network of <figref idrefs="f0006">figure 6</figref> in an analysis phase,</dd><dt>Fig. 9</dt><dd>one on the image of <figref idrefs="f0004">figure 4</figref> Returning result image of a direction-independent gray value analysis,</dd><dt>Fig. 10</dt><dd>one on the image of <figref idrefs="f0004">figure 4</figref> Returning result image of a direction-dependent gray value analysis,</dd><dt>Fig. 11</dt><dd>from deciding whether the pixels belong to the swath or the field, going out image from the image <figref idrefs="f0007">figure 9</figref> back, </dd><dt>Fig. 12</dt><dd>the image of <figref idrefs="f0008">figure 11</figref> after removal of all areas which are smaller than 450 pixels,</dd><dt>Fig. 13</dt><dd>the remaining, largest area of ​​the image in <figref idrefs="f0008">Fig. 12</figref>and</dd><dt>Fig. 14</dt><dd>the end result of the processing of the image from <figref idrefs="f0009">Fig. 13</figref> with marked inertia axes.</dd></dl>

p0030In the <figref idrefs="f0001">figure 1</figref> is shown a harvesting machine in the form of a self-propelled forage harvester 10th The forage harvester 10 is built on a frame 12 that is carried by front driven wheels 14 and steerable rear wheels worn 16th The operation of the forage harvester 10 is controlled from an operator's cab 18 from which is visible from a harvested crop 20th By means of the crop pickup 20 picked up from the ground, z. B. grass or the like is supplied through non-illustrated feed rollers, which are arranged within a feed housing on the front side of the forage harvester 10, a chopper drum 22 that chops it into small pieces and a conveyor 24 gives up. The crop leaves the forage harvester 10 to accompanying trailer over a about a substantially vertical axis rotatable and adjustable in the tilt discharge duct 26. Between the chopper drum 22 and the conveyor 24 extends a post-chopper reduction 28 through which the material to be conveyed to the conveyor 24 is fed tangentially.

p0031The crop pickup 20 is formed in this embodiment as a so-called pick-up. The crop pickup 20 is built on a frame 32 and is supported mounted on both sides support wheels 38 which are mounted on a respective support 46 to the frame 32, on the ground from. The object of the crop pickup device 20 is incorporated on the ground of a field in a swath 48 Undocked crop and 10 feed the forage harvester for further processing. To this end, the crop pick-up 20 is moved during the harvesting operation with a small distance to the ground via the field while it is raised for transport on a road or on paths. For the crop pickup 20 includes a conveyor 36 in the form of a screw conveyor 20 conveys the crop picked up from the sides of the harvested crop to a located in the center, not shown dispensing opening, behind the follow the feed rollers. The Erntegutvorrichtung 20 also includes a, as well as the conveyor device 36, driven in rotation pickup 34, which is arranged below the conveyor 36 with its conveyor prongs the good from the ground raises to hand it over to the conveyor 36th In addition, a hold-down device 40 is fixed in the form of an over the transducer 34 on the frame plate 32nd

p0032The forage harvester 10 is fitted at the top of the leading in the direction of travel side of the cab 18 with a camera 42nd The camera lens 42 is directed obliquely to the swath 48 forwards and downwards. The camera 42 is located on the longitudinal central axis of the forage harvester 10. The camera 42 forms with an electronic controller, a below-described automatic steering system that guides the forage harvester 10 automatically along the swath 48, to facilitate the work of the operator in the cab 18th

p0033First, however, is based on the <figref idrefs="f0002">figure 2</figref> another possible application of the automatic steering system described. Here, the camera 42 is mounted at the top of the leading side in the direction of travel of the operator's cab 18 of a tractor 50th It is located on the longitudinal center plane of the tractor and its lens is directed to the windrow 48 is also forwardly and downwardly. The tractor 50 has front steerable wheels and rear driven wheels 54th He pulls her behind a baler 56, which receives the crop out of the swath 48 by means of a transducer 34 from the field and formed into bales 58th Instead of the round baler can also be a rectangular baler baler 56th It could also be designed to be self-propelled. The invention can also be used on a serving for tillage tractor 50th

p0034The forage harvester 10 and the tractor 50 are fitted on their front sides with lighting devices, which enable the swath illuminate 48, if the available light is not sufficient.

p0035In the <figref idrefs="f0003">figure 3</figref> is the steering system 60 of the forage harvester 10 from <figref idrefs="f0001">figure 1</figref> or the tractor 50 from <figref idrefs="f0002">figure 2</figref> shown schematically. The camera 42 includes an image sensor 62, in particular in the form of a CCD or CMOS to digitized the output signal using an analog-to-digital converter 64 and a video interface 66 is supplied. The image data from the camera 42 may include only luminance (gray values, brightness) exist or chrominance (color). A processor 68 (usually microprocessor or controllers) communicates bi-directionally with the video interface 66. In a related processor 68 display device 70, the image 42 produced by the lens 82 of the camera on the image sensor 62 can be displayed to the operator in the cab 18 , The display device 70 can display other information, such as operational data of the steering system 60 and / or of the forage harvester 10 or the tractor 50th The operator is also a keyboard 72 or other input device (eg. A mouse, joystick, speech recognition) is available, with which he can perform inputs to the processor 68. The processor 68 controls electromagnetic valves 74, the input side under pressure with a source 76 hydraulic fluid and the output side are connected to a steering cylinder 78th The steering cylinder 78 is adjusted in the forage harvester 10 from<figref idrefs="f0001">figure 1</figref> the rear wheels 16 and the tractor 50 from <figref idrefs="f0002">figure 2</figref> the front wheels 52. A wheel angle sensor 80 detects the current steering angle of the wheels 16 and 52, and supplies this information to the processor 68 to. The hardware configuration of such automatic steering systems 60 is well known in the art, so that it can be omitted here a more detailed description and referring to the mentioned in the introduction to the art, the disclosure of which is incorporated by reference present in the documents.

p0036The task of the automatic steering system 60 is 50 to guide the forage harvester 10 or the tractor with the baler 56 with no steering action of the operator along the swath 48 on the field. The processor 68 processes to the images of the camera 42 by electronic image processing and generated from the results appropriate control signals for the solenoid valves 74th

p0037The swath 48 is usually made of dried stalks, which cut with a mower, left lying on the field for drying and merged by Rake or when harvesting a cornfield is filed through the combine in the swath. A typical, locked in the grass crop by the camera 42 image is in the<figref idrefs="f0004">figure 4</figref> reproduced. It is apparent that differ brightness values ​​between the swath 48 and the areas of the field 84 laterally adjacent to the swath 48 is not very strong. It would therefore be difficult or almost impossible to generate a steering signal by the individual pixels (hereinafter pixels) 42 assigns the swath 48 and the field 84 of the image signal of the camera only by their brightness. A chrominance data containing image could then be more appropriate to distinguish between the swath 48 and Section 84 on the basis of color, when the stalks in color differs sufficiently from the field. But this is especially not the case, though standing on the field remaining blades are at least partly dried, or if the swath is to be taken immediately after mowing, as in the grain harvest.

p0038Based on <figref idrefs="f0004">figure 4</figref> but is also recognized that in the field 84 essentially only vertically standing stalks (grass stubble) are present, while the swath 48 in contrast, includes long and lying in different, to some extent horizontally extending orientations straws. The present invention therefore proposes to evaluate the texture of the pixels and for determining whether a pixel in the field 84 or the swath 48 is to consult. Under Texture is the local distribution and variation of the gray scale values ​​(and / or chrominance) substance in a portion of the image. On the field, the texture of short, mostly vertical structures is (stems), while it is in the swath of longer blades which are in different directions.

p0039Accordingly, the processor 68 according to a procedure as in the works <figref idrefs="f0005">figure 5</figref> is shown. Starting from an image taken in step 100 image of the camera 42, is provided via the video interface 66 to the processor 68 a pixel file available in step 102nd From this pixel file a respective texture information is derived in step 104 for processed further pixel. In step 104, all incoming pixel or only some of them containing a the swath 48 containing part of the image will be processed to reduce the computation time.

p0040The derivation of texture information in step 104 can be carried out in various ways. It is possible to perform a gray value depending analysis, in which a dependency matrix is calculated. The dependency matrix contains information which combinations of adjacent gray values ​​occur in a neighborhood of pixels for small image areas. Are the gray values ​​in the neighborhood of the pixel homogeneous, therefore, the dependency matrix out that only identical gray values ​​are present. Varying the gray values ​​in the neighborhood, however very strong, another Depending gives stiffness matrix. In a preferred embodiment, only the direct neighbors of the eight pixels are considered in the gray value analysis, on the edge of the evaluated image area reduced their number to three (in the corners) or five. Is obtained in this way, information as to whether the texture in the neighborhood of the pixel is uniform or not.

p0041The derivation of texture information in step 104 may also be effected by a direction-dependent gray value dependency analysis. As seen from<figref idrefs="f0004">figure 4</figref> is recognized and has already been mentioned above, the field 84 contains a number of vertical line structures. In swath 48, however, the blades are very confused. It is therefore advisable to allow the diagonal neighbors of pixels taken into account and to use only the pixels above and below as well as the left and right of the subject pixel to create a dependency matrix. The deduced in this way in step 104 texture information thus contains information as to whether the texture in the vicinity of the pixel contains vertical or horizontal structures. The amount of information, but not necessarily, the information content is reduced compared to the non-directional gray value dependency analysis.

p0042Further, the texture information may be generated in step 104 by a color analysis. In this study, the chrominance information of the swath 48 is analyzed. If a colored working camera 42 used each pixel of the generated color image has three color information, each for red, green and blue. If the swath 48 so different in color from the box 84 may be inferred from the color information easily between the two, even with the same brightness, can be distinguished. Finally, it is also possible to combine the gray level dependency analysis or the direction-dependent gray value analysis dependence with the color analysis. omnidirectional or directional dependency matrix for the gray values ​​and color information It is generated for each evaluated pixels that are processed together. In the said analyzes particular comes the so-called second order statistics are used. In addition, a series combination is possible in which first carried out a pure RGB-analysis, and then the result of the RGB-analysis is subjected to a texture analysis.

p0043After information about the texture in the vicinity of the pixel obtained in step 104 follows according to the <figref idrefs="f0005">figure 5</figref> step 106. There is carried out based on the texture information provided in step 104, a classification, ie decided for each pixel to be processed or pixel group, whether it belongs to the swath or field. Here any suitable algorithms can be used. Be particularly advantageous to use a neural network has been proven as it can adapt itself for its ability to learn also changed image recording conditions. Another possibility would be the use of a fuzzy logic or neuro-fuzzy logic. This decision is made in two stages: first a probability value is evaluated for each pixel or each pixel group, which represents the probability of it or it belongs to the swath. Subsequently, it is checked whether the respective probability value (50% usually) is above a threshold or not.

p0044A possible embodiment of such a neural network 112 is in the <figref idrefs="f0006">figure 6</figref> shown schematically. The network 112 is composed of at least two layers of neurons 114, 116th The first neuron layer 114 has an input 118 to which the texture information obtained in step 104 is supplied in the form of the dependency matrix and optionally the color information. In the neuron layer 114 a combination of the input values ​​with learnable information whose result at an output 120 of the first neuron layer 114 is ready and the input 122 of the second neuron layer 116 is supplied takes place.

p0045As shown, the first neuron layer 114 of the second neuron layer 116 multiple outputs available in parallel. The second neuron layer 116 has a single output 124. Also in the second neuron layer 116 a link is made to the present at its input 122 signals learnable information. At the output of the second neuron layer 116 to a binärisierende information about is finally provided, if the examined pixel or pixel group is part of the swath 48 or the field 84th The network 112 may also include additional layers of neurons between the two illustrated neuron layers 114, 116th It can also be any other network architecture than the illustrated, so-called back-propagation network can be used.

p0046For each pixel to be examined its own neural network 112 could be provided, or there is a single network 112 used the input data are supplied to all pixels sequentially. Typically, the neural network 112 is implemented in software by the processor 68th In other embodiments, however, it could also be realized by dedicated hardware.

p0047As mentioned above, the neural network 112 is capable of learning. He is thus initially taught which parts of a captured image to the swath 48 and which belong to the field 84th This process is in the<figref idrefs="f0006">figure 7</figref> shown schematically. The neural network is supplied (instead of from an image that was captured by the camera 42, derived texture information) which generated from a learning image texture information, which can be, for example, stored electronically, or the forage harvester 10 or the tractor 50 is in the field 84 positioned in front of a swath 48 and the swath 48 is recorded with the camera 42nd Furthermore, an information on the position of areas of the swath 48 and the array 84 is supplied to the respective image to the neural network 112th It may be rectangles 86, 88, as in the<figref idrefs="f0004">figure 4</figref> are shown. It masks are thus defined, are contained in the interior of the box 84 and the swath 48 belonging pixel quantities. The neural network 112 is based on this information to learn, able to as the swath 48 and the field look 84 and to distinguish them. This learning process could also be done by the driver controls the forage harvester 10 or the tractor 50 by hand along a swath 48th The neural network 112 learns also to distinguish between the swath 48 and field 84th The neural network 112 delivers a result image that is used to success examination. It can be reproduced on the display device 70 and inform the operator whether the working conditions for the automatic steering system 60 are sufficiently or if, should be better accessed for example in darkness or fog to manual steering. The learning phase of the neural network 112 can be repeated repeatedly a corresponding operator indication toward or extended, or otherwise as described above, be carried out once during the manufacture of the steering system 60 or it may be provided for changing conditions selectable memory values ​​or removable memory cards.

p0048After the learning phase (<figref idrefs="f0006">figure 7</figref>) Is completed, the neural network 112 operates in accordance with the <figref idrefs="f0006">figure 8</figref>, He is from images from the camera 42 derived texture information supplied and it delivers a pixel image in which there is a binary distinction between swath and field. By subjecting the neural network 112 with the texture information, the distinction between the swath 48 and Section 84 is easier and possible even at low visibility conditions.

p0049Referring now again to the <figref idrefs="f0005">figure 5</figref> Referring, in the following step 106 now exists a binary bitmap file. The individual pixels are associated with either the field 84 or the swath 48th Based on these pixel file the longitudinal axis of the swath 48 is determined in step 108, ie it is determined their direction (angle to the longitudinal central axis of the forage harvester 10 or the tractor 50) and their distance from the longitudinal central axis of the forage harvester 10 or the tractor 50th a steering signal is determined from the direction and the distance in step 110 then generated which is supplied to the electromagnetic valves 74th The forage harvester 10 or the tractor 50 moves thus automatically along the swath 48. The in<figref idrefs="f0005">figure 5</figref> illustrated procedure is repeated regularly, for example, 25 times a second. It is also higher or lower repetition frequencies are used. At the end of the swath, the operator steers the harvester manually to the next male swath. In another embodiment, the next swath is recognized by the automatic steering system 60 and the harvester directed there automatically.

p0050The <figref idrefs="f0007">figures 9</figref> to 15 show illustrative images in during <figref idrefs="f0005">figure 5</figref> processing shown in the in the <figref idrefs="f0004">figure 4</figref> emerge image shown.

p0051The <figref idrefs="f0007">figure 9</figref> shows a result for a non-directional gray value dependency analysis. The<figref idrefs="f0007">figure 10</figref> shows a result for a direction-dependent gray value dependency analysis. In both figures, calculated by the neural network probability values ​​are shown that the pixel belongs to the swathe before being binarized. A comparison of the<figref idrefs="f0007">Figures 9 and 10</figref>, Both quasi the result of step 106 of <figref idrefs="f0005">figure 5</figref> pose before the classification, shows the positive influence of the direction-dependent gray value dependency analysis.

p0052For further processing, therefore the results of the direction-dependent gray value dependency analysis will only be used. The<figref idrefs="f0008">figure 11</figref> represents the result of the decision (step 106 in <figref idrefs="f0005">figure 5</figref>) Represents whether the pixel to the swath 48 (white) or field 84 (black) belong, ie the binarization by the neural network 112. The neural network 112 has previously been subjected to a learning phase (<figref idrefs="f0006">figure 7</figref>), In which he made the image <figref idrefs="f0004">figure 4</figref> as a learning image and the rectangles 86, 88 as to the field 84 or 48 belonging swath areas were notified. This network also showed other images good results.

p0053The <figref idrefs="f0008 f0009">Figures 12 to 14</figref> corresponding process steps of step 108 of <figref idrefs="f0005">figure 5</figref>, In the<figref idrefs="f0008">figure 12</figref> all contiguous areas were calculated and all faces away that are smaller than a threshold value (here 450 pixels). at<figref idrefs="f0009">figure 13</figref> is the largest area of ​​the <figref idrefs="f0008">figure 12</figref> been left over, then when in <figref idrefs="f0009">figure 14</figref> the little black interfering surfaces were removed by use of dilation and erosion. In the<figref idrefs="f0009">figure 14</figref> are also the two axes of inertia of the remaining area of ​​the swath 48 located. With the extending longitudinally axis of inertia, the direction of the swath 48 and its distance from the longitudinal center plane of the forage harvester 10 or the tractor 50 can be determined. These quantities are used to generate a steering signal for the electromagnetic valves 74, wherein the current value of the wheel angle sensor 80 is taken into account.