Mesh-based processing of measurement data for classifying and determining attributes of objects in region

Abstract

A method for processing measurement data from monitoring of an area into classification scores and / or attributes of objects in the area with respect to a given classification, comprising: providing a point cloud of measurement data; providing an input grid having a plurality of cells; assigning a value of the at least one measurement quantity and / or a value of at least one work product derived from the at least one measurement quantity to each cell of the input grid based on the point cloud; and processing the input grid into an output grid by a task neural network, each cell being determined to be o occupied by an object if at least one point of the point cloud is within the cell, and o unoccupied if the cell does not have points of the point cloud; and the value assigned to the unoccupied cells in the respective grid is modified to a neutral value in the context of the respective grid.

Description

Technical Field

[0001] The invention relates to the processing of measurement data in the form of point clouds that have been collected by monitoring an area. The results of the processing include classification scores and / or other properties of objects in the area. Background Art

[0002] In many applications where an area or volume in a space is monitored with the aid of one or more sensors, a representation of the area including information about the objects in the area is required. Specifically, it is necessary to determine which parts of the monitored area are occupied by which types of objects, and what other properties these objects have.

[0003] Some physical modalities used to monitor the area (such as radio radar and lidar) produce measurement data in the form of point clouds. The point cloud assigns a value of measurement data to each point contained therein. These points can be anywhere in space and do not need to form a regular grid, or even a continuous area. The measurement data is projected onto such a regular grid and then further processed into the classification score and / or other attributes sought. Summary of the invention

[0004] The present invention provides a method for processing measurement data from monitoring an area into information about objects in the area. The information includes classification results for a given classification, and / or other sought properties of the objects.

[0005] The method starts with a point cloud of measurement data. The point cloud assigns a measurement value of at least one quantity to each of a plurality of points in space. The points may be specified by any suitable set of coordinates, such as Cartesian or polar coordinates. The coordinate system is typically determined by the type of measurement being performed. For example, in a radio radar measurement, a point may be characterized by a distance, an azimuth angle, and an elevation angle. Specifically, each point in the point cloud may represent a location on an object from which the signal causing the measurement value originated.

[0006] A point cloud is a list of unordered points, i.e. the order is irrelevant. Each point is characterized by certain properties, such as its radar cross section (RCS) or radial velocity compensated by the radar's own motion.

[0007] In addition, an input grid having a plurality of cells is provided. Each cell represents a specific volume in space. The resolution with which the area is monitored determines the volume of each cell. The grid may be defined as a three-dimensional grid or a two-dimensional grid. Specifically, a two-dimensional grid defined in a plane may be considered as a three-dimensional grid having cells extending from the plane to infinity in a direction perpendicular to the plane.

[0008] Based on the point cloud, a value of at least one measured quantity and / or a value of at least one work product derived therefrom is assigned to each cell of the input grid. This can be considered as a projection of the point cloud onto the input grid. The projection can be performed in any suitable way. For example, the measured values ​​attached to the points in each cell can be aggregated into one value, which is then assigned to the cell. There are many possibilities for doing this. For example, the measured values ​​can be added and the sum can be divided by the number of contributing points. Depending on the position of the point in the cell, the measured values ​​can also be weighted differently in the sum.

[0009] The task neural network processes the input grid into an output grid, the cells of which carry the classification scores and / or attributes of the objects. The output grid does not have to be the same as the input grid. In particular, the output grid may be coarser than the input grid, that is, it may contain fewer cells. The classification scores and / or attributes of the objects also imply assumptions about whether the object exists in a specific area corresponding to the cells of the output grid, or whether there are no objects in the cells of the output grid. The task neural network can generate the output grid in one go, but it can also generate intermediate grids in the process during the processing of the input grid into the output grid. In particular, if the architecture of the task neural network consists of two different parts, the first part can generate an intermediate grid, and the second part can continue to process the intermediate grid into the output grid.

[0010] Specifically, the attributes of the object may include one or more attributes of the oriented bounding box OBB that characterizes the object. Examples of these attributes include existence probability, size, orientation, and object type. That is, when predicting the OBB, classification (here: about object type) and regression (here: about size and orientation) may be combined. The existence probability may also be determined as a numerical value by regression. Alternatively, the OBB may be classified in a binary manner as to whether it is likely to exist.

[0011] During the course of the method, the input mesh, the output mesh and / or intermediate meshes occurring during the processing of the input mesh into the output mesh are modified.

[0012] For each cell of the corresponding grid, based on the coordinates of the points in the point cloud, determine which cells are occupied by an object and which cells are not occupied by an object. If at least one point of the point cloud is within a cell, then the cell is determined to be occupied by an object. If the cell has no points in the point cloud, then the cell is considered unoccupied. In other words, if no signal is emitted from an object anywhere within the cell that causes a measurement value, then the cell can be considered to have no object.

[0013] To modify the corresponding grid, the values ​​assigned to the unoccupied cells in the corresponding grid are modified to neutral values ​​in the context of the corresponding grid.

[0014] It has been found that various stages of processing may result in one or more cells being assigned classification scores or other properties of an object that are not completely consistent with the absence of an object, even if there is actually no object in the cell. For example, if the projection of a point cloud of measurement data onto an input grid is performed by a neural network, small but non-zero values ​​of at least one measurement quantity may be assigned to input grid cells where there is actually no object. During subsequent processing into an output grid, these values ​​may be amplified and cause unoccupied cells in the output grid to be assigned classification scores and / or properties that indicate the presence of an object where there is no object. For example, if the obtained classification scores and / or properties of the objects are assembled into a representation of the vehicle environment, "phantom" objects may appear. The vehicle's trajectory planning or another vehicle assistance system may then react to these "phantom" objects in an inappropriate manner. For example, the automated vehicle may brake suddenly or initiate an evasive maneuver to avoid a collision with a "phantom" object.

[0015] Likewise, when a task neural network processes an input grid into an output grid, unwanted "filling in" of unoccupied grid cells with information that might be perceived as indicating the presence of an object can also occur. Intermediate grids computed in the process can also be affected by this "filling in."

[0016] By modifying the input grid, the intermediate grids, and / or the output grid to include neutral values ​​for unoccupied cells, the corresponding grids are adjusted to be consistent with the knowledge from the point cloud about where the object is located in space. That is, the measurements from the point cloud are further processed and enhanced until the sought classification score and / or properties of the object are obtained, but the initially acquired knowledge about the location of the object in space remains intact. The presence or absence of a signal emanating from an object in a particular cell is considered to be the most reliable indicator of the presence or absence of the object.

[0017] In this article, which values ​​of the classification score and / or other attributes can be considered as neutral values ​​depends on the specific application at hand. For example, if the classification scores are all related to certain types of objects, a full zero classification score can be a neutral value indicating that there is no object. However, if there is another dedicated category for a background without objects, the one-hot classification score for that background category can be considered as a neutral value. In another simple example where the final derived attribute includes the temperature of the object, the neutral value is not the freezing point or even absolute zero. Instead, the temperature of the object-free unit should be merged with the ambient temperature.

[0018] Therefore, in a particularly advantageous embodiment, a value relating to an object-free background in the context of the respective grid is selected as a neutral value.

[0019] Furthermore, what value is considered a neutral value may also depend on the loss function (objective function) used as a source of feedback during the training of the task neural network.

[0020] Modifying the values ​​for unoccupied cells to neutral values ​​can be called "occupancy masking". In this way, the neural network is focused only on occupied cells (containing individual points) during training. In addition, the neural network can only output bounding boxes in occupied cells. This can be beneficial in cases where the point cloud is sparse. For example, when a radio radar sensor outputs a low number of points or when considering objects that are far away.

[0021] Occupancy masking does not modify the network architecture. It only modifies the output class probabilities and regression parameters. This is true during both training and inference. Therefore, it affects both the model behavior during model training and inference.

[0022] Occupancy masking takes as input the output mesh of the object hypothesis and masks the mesh based on the input point cloud. The information in occupied cells remains the same, while the information in unoccupied cells is masked.

[0023] The abstract features extracted from the point cloud can then be processed using a backbone convolutional neural network that processes data at multiple spatial scales. For a specific implementation, a backbone consisting of a residual network and a feature pyramid network can be used, which is capable of extracting features at different output resolutions of the 2D grid in the backbone.

[0024] Because objects may span several grid cells, the generated object hypotheses can be filtered by voting. This is done by non-maximum suppression. For each object, the spatially superimposed object hypotheses can be filtered according to the object with the highest object probability. The filtered object hypotheses in the form of OBB are then the output of the detection network.

[0025] Advantages of the approach proposed here include:

[0026] Unoccupied cells do not produce predicted bounding boxes because their outputs are masked by the occupancy mask. This forces detections to only exist at occupied locations (the model does not make random guesses in empty space).

[0027] Since the network training only takes into account occupied cells, better detection performance is expected.

[0028] It performs better when the point cloud is sparse (such as the output of a radio radar sensor) or when working at long distances.

[0029] Occupancy masking reduces the complexity of post-processing steps in the object detection pipeline. In effect, it removes bounding box predictions from unoccupied grid cells and thereby reduces the number of predicted bounding boxes. Non-maximum suppression is typically applied to remove duplicate boxes and scales with the number of boxes. Hence, its complexity is greatly reduced.

[0030] The method offers significant benefits, has limited disadvantages, and can be used for a wide range of products.

[0031] Where the occupancy masking is best applied depends on the specific application and in particular on what needs to be guaranteed in such application. For example, applying the occupancy masking at the output of the task neural network guarantees that the output will have no objects in unoccupied cells. On the other hand, applying the occupancy masking as early as possible after the initial mesh projection reduces the computational complexity of further calculations.

[0032] In yet another advantageous embodiment, in addition to the categories associated with the objects, the given classification also includes categories associated with the object-free background. In this way, the absence of an object can be distinguished from other situations where, for example, the classification result is completely uncertain or may result in an all-zero classification score. For example, if a completely new type of object appears in the scene, but the given classification only includes a limited set of categories associated with known objects, then both this completely new object and the object-free background may receive an all-zero classification score.

[0033] In yet another particularly advantageous embodiment, the value assigned to the unoccupied cell is further modified according to the value in one or more neighboring cells. In this way, artifacts produced by hard boundaries between values ​​that remain unchanged and values ​​that are modified to neutral values ​​can be suppressed. This further modification can also be considered "filling".

[0034] For example, a softening algorithm may be applied to at least one boundary between a first region including occupied cells and a second region including unoccupied cells. An exemplary way of doing this is to apply max pooling to a grid having a kernel size of, for example, 3. This is equivalent to setting all empty cells adjacent to occupied cells (including diagonals) as occupied cells. In another example, in response to determining that a cell is occupied, adjacent cells may also be considered occupied.

[0035] The padding operation can be applied to either the input mesh or the output mesh. The input mesh is the finest scale of the network after mesh projection of the point cloud. The output mesh is the scale at which the detection head operates, which is usually coarser than the scale of the input mesh. The output mesh is obtained by shrinking the input mesh to the scale of the corresponding detection head. Experiments show that output padding with a kernel size of 3 significantly improves performance across settings.

[0036] In theory, the input mesh is not necessary, since the output mesh can be computed directly from the point cloud. However, it is computationally efficient to compute the output mesh from the input mesh. In practice, once the point cloud is projected to a mesh, the input mesh is available anyway. This is why it may be advantageous to have padding on both the input occupancy mesh and the output occupancy mesh.

[0037] In yet another particularly advantageous embodiment, the values ​​in the intermediate grid and / or the output grid are selected to represent logits of a precursor to a classification score. In this way, if the network is configured to perform the final processing of converting logits into classification scores differently during training and during inference (i.e., using a trained network), the modification to avoid "ghost" objects only needs to be applied once and can then be used both during training and during inference. For example, a network that uses a non-linear, non-differentiable activation function to determine the final output of a neuron during inference can use a differentiable approximation of that activation function during training to allow back-propagation of the gradient of a loss function (objective function) used to optimize the network parameters.

[0038] In a particularly advantageous embodiment, during the training of the task neural network, the calculation of the classification scores for the elements is merged with the calculation of the loss function values ​​for ranking these classification scores. This avoids numerical instabilities due to very small or very large values. That is, instead of calculating very small or very large classification scores and then inserting them into the calculation of the loss function, the classification scores and the loss function values ​​are calculated at once. The merged calculation is an example of how the classification scores can be calculated differently during training on the one hand and during inference on the other hand.

[0039] Typically, the class probabilities are obtained by applying a softmax operation to the unregularized output of each unit of the neural network (also called a pairwise unit). For example, when the pairwise unit of a grid cell is [1, 0, 0], the corresponding probability is softmax([1, 0, 0]) = [0.57, 0.21, 0.21].

[0040] Commonly used loss functions for classification (such as cross entropy and focal loss) are usually applied directly to pairs rather than probabilities after softmax. This makes the formulation of these loss functions more numerically stable.

[0041] This is why the new dedicated method of setting class probabilities in unoccupied cells is advantageous. This leads to a simpler implementation by supporting a shared implementation between training and inference. In fact, training directly optimizes the loss function output and does not require class probabilities as direct output. On the other hand, inference requires class probabilities computed from the meta-elements.

[0042] Thus, in yet another particularly advantageous embodiment, for unoccupied cells, the background pair may be set to an arbitrarily high value, such as 10, and the other pairs may be set to arbitrarily low values, such as 0. Any high / low value combination with a sufficiently high difference is usable.

[0043] In yet another advantageous embodiment, a maximum of 10% of the cells in the grid are occupied. That is, the grid is only sparsely occupied. If the point cloud is sparse, calculating the final classification score and / or the properties of the object in the presently proposed manner is more stable and more reliable than a calculation that does not modify the values ​​of the unoccupied grid cells. For example, a task neural network for object detection may output bounding boxes only in occupied cells. Typically, when monitoring a vehicle environment with a radio radar or lidar sensor, there are several thousand points associated with the object, which is approximately one hundred thousand grid cells. Exemplary cases where the point cloud is particularly sparse include cases where the radio radar sensor outputs only a low number of points or is monitoring distant objects.

[0044] In yet another particularly advantageous embodiment, the task neural network comprises

[0045] ● a feature extractor or object detector that produces a feature map or bounding box as the intermediate grid, and

[0046] ● At least one classification head that further processes these intermediate grids into output grids.

[0047] In this way, the same intermediate grid can be reused by several classification heads for further processing. For example, different classification heads may be responsible for detecting different types of objects. For example, when monitoring a vehicle's environment, one classification head may be responsible for classifying traffic signs, while another classification head may be responsible for classifying other traffic participants. Therefore, a classification head may also be called a "category head".

[0048] In the classification head, an additional convolutional neural network estimates the object probability between 0 and 1 for each cell of the feature map and the regression parameters (position, length, width, height, orientation) of the oriented bounding box OBB. Several of these classification heads are used for the detection of different object types, each of which is responsible for the prediction of the object type category (i.e., object types with similar properties, such as trucks and buses). These classification heads use feature maps with appropriate resolution depending on the type of object to be detected. For example, feature maps with higher resolution are used for small objects (such as pedestrians) rather than for large objects (such as trucks).

[0049] The network outputs one or more grids of object hypotheses (usually one grid per class head). For each cell, one or more object hypotheses are predicted with object type classification, location of the object, and dimensions (length, width, height, orientation) of the OBB surrounding the object.

[0050] In yet another particularly advantageous embodiment, the measurement data comprises radio radar data, lidar data and / or ultrasound data acquired using at least one sensor. These are prime examples of measurement modalities that produce sparsely distributed point clouds.

[0051] In yet another particularly advantageous embodiment, a projection neural network is used to assign the value of at least one measured quantity and / or the value of at least one work product derived therefrom to each cell of the input grid. This approach produces fewer artifacts than simply directly calculating the value of the measured quantity to be assigned to each cell of the input grid from the points in that cell.

[0052] The input of the grid projection module is unordered points. Its output is an n-dimensional (usually 2 or 3-dimensional) regular grid covering a specific area. The resolution determines the area of ​​each of the multiple cells in the grid.

[0053] Mesh projection can also be done in a multi-scale manner, where the input is a list of unordered points and the output is multiple n-dimensional regular grids. Typically, each regular grid has a corresponding grid of the same spatial dimension in the backbone network, so that they can be fused in any way (concatenation, addition, etc.).

[0054] In yet another particularly advantageous embodiment, an actuation signal is calculated based on the determined classification score and / or attribute of the at least one object. The actuation signal is used to actuate a vehicle, a traffic assistance system, a robot, a surveillance system, a quality inspection system and / or a pick-and-place manufacturing system. In this way, the probability is increased that the action performed by the corresponding actuated system in response to the actuation signal is appropriate in the case characterized by the raw measurement data.

[0055] For example, in automated manufacturing or assembly systems, based on the final calculated classification scores and / or attributes, components and their orientation can be detected to determine their gripping points. Based on the final calculated classification scores and / or attributes, vehicles and robots (such as lawn mowers) can detect obstacles, traffic signs, other traffic participants that have an impact on the future behavior of the vehicle or robot. With necessary modifications, this also applies to traffic assistance systems, in particular for bicycles or other two-wheeled vehicles. However, traffic applications are not limited to on-board sensors. On the contrary, traffic can also be monitored, for example, using fixed radio radar or lidar sensors. Monitoring systems or access control systems can detect and identify authorized persons based on the final calculated classification scores and / or attributes so that alarms can be disabled and / or doors can be opened automatically. Similarly, monitoring systems can also detect, classify and identify dangerous goods.

[0056] The method may be implemented in whole or in part by a computer and embodied in software. Therefore, the present invention also relates to a computer program having machine-readable instructions, which, when executed by one or more computers and / or computing instances, cause one or more computers and / or computing instances to perform the method described above. In this document, control units for vehicles or robots and other embedded systems capable of executing machine-readable instructions are also considered computers. Computing instances include virtual machines, containers, or other execution environments that allow machine-readable instructions to be executed in the cloud.

[0057] The non-transitory machine-readable data carrier and / or the download product may include a computer program. The download product is an electronic product that can be sold online and transmitted over a network for immediate fulfillment. One or more computers and / or computing instances may be equipped with the computer program and / or equipped with the non-transitory storage medium and / or the download product.

[0058] Hereinafter, the present invention is described using the accompanying drawings, which are by no means intended to limit the scope of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0059] The accompanying drawings show:

[0060] Figure 1 : An exemplary embodiment of a method 100 for processing measurement data 3 from monitoring an area 1 having an object 2;

[0061] Figure 2 : Illustration of occupancy masking performed during method 100. DETAILED DESCRIPTION

[0062] Figure 1is a schematic flow chart of an embodiment of a method 100 for processing measurement data 3 from monitoring an area 1 with objects 2. The results of the processing include classification scores 2a and / or attributes 2b of the objects 2 in the area 1 with respect to a given classification.

[0063] In step 110, a point cloud 3* of measurement data 3 is provided. The point cloud 3* assigns a measurement value of at least one quantity to each of a plurality of points 3a-3e in space.

[0064] In step 120, an input grid 4 having a plurality of cells 4a is provided.

[0065] In a step 130 , based on the point cloud 3 *, a value 4 b of at least one measured quantity and / or at least one work product derived therefrom is assigned to each cell 4 a of the input grid 4 .

[0066] In step 140 , the task neural network 5 processes the input grid 4 into an output grid 6 , the cells 6 a of which carry the classification scores 2 a and / or attributes 2 b of the objects 2 as values ​​6 b .

[0067] According to blocks 131 and / or 141, each cell 4a, 6a, 7a of the input grid 4, the output grid 6 and / or of the intermediate grid 7 occurring during the processing of the input grid 4 into the output grid 6 is determined as:

[0068] if at least one point 3a-3e of the point cloud 3* is within the cell 4a, 6a, 7a, then it is occupied by the object 2, and

[0069] If the cell 4a, 6a, 7a has no point 3a-3e of the point cloud 3*, it is unoccupied U.

[0070] According to blocks 132 , 142 , respectively, the values ​​4b , 6b , 7b assigned to the unoccupied cells 4a , 6a , 7a in the respective grids 4 , 6 , 7 are then modified to neutral values ​​in the context of the respective grids 4 , 6 , 7 .

[0071] According to blocks 132a, 142a, respectively, in the context of the corresponding grids 4, 6, 7, the value associated with the object-free background may be selected as a neutral value.

[0072] According to blocks 132b, 142b, respectively, the value 4b, 6b, 7b assigned to an unoccupied cell 4a, 6a, 7a may be further modified according to values ​​in one or more neighboring cells 4a, 6a, 7a.

[0073] According to block 133 , a value 4 b of the at least one measured quantity and / or at least one work product derived therefrom may be assigned to each cell of the input grid using a projection neural network 8 .

[0074] According to block 143 , in addition to the category associated with object 2 , the given classification may include a category associated with a no-object background.

[0075] According to block 144, a value in the intermediate grid 7 and / or the output grid 6 may be selected to represent a pair that is a precursor to the classification score 2a.

[0076] According to block 145, during the training of the task neural network 5, the calculation of the classification scores 2a based on the meta-elements can be combined with the calculation of the loss function values ​​for ranking these classification scores 2a.

[0077] According to block 146, the task neural network 5 may include

[0078] A feature extractor or object detector 5a, producing a feature map or bounding box as an intermediate grid 7, and

[0079] ● At least one sorting head 5b to further process these intermediate grids 7 into output grids 6.

[0080] exist Figure 1 In the example shown in FIG. 1 , in step 150, an actuation signal 150a is calculated based on the determined classification score 2a and / or attribute 2b of at least one object 2. In step 160, the vehicle 50, the traffic assistance system 51, the robot 60, the monitoring system 70, the quality inspection system 80 and / or the pick and place manufacturing system 90 is then actuated using the actuation signal 150a.

[0081] Figure 2 Occupancy masking performed during method 100 for a simple example is illustrated.

[0082] In this example, the point cloud 3* includes five points 3a-3e. In step 130 of the method 100, these five points 3a-3e are projected into a given input cloud 4, with the unit 4a as the value 4b. Optionally, according to block 133 of the method 100, this step can be performed by the projection neural network 8.

[0083] The task neural network 5 includes an object detector 5a, which processes the input grid 4 into an intermediate grid 7 with cells 7a and values ​​7b, which represent the bounding box of the object 2. The task neural network 5 also includes a classifier head 5b, which processes the intermediate grid 7 into an output grid 6 with cells 6a and values ​​6b. These values ​​6b represent the sought classification scores 2a and / or attributes 2b of the objects 2 in the observed area 1. Figure 2 In the simple example shown in , the output grid 6 has the same spacing as the input grid 4 .

[0084] exist Figure 2In the example shown in , the output grid 6 is modified by occupying the mask according to blocks 141 and 142 as described above. This means that for the positions in the output grid 6 corresponding to the points 3a to 3e, the values ​​6b in the output grid 6 remain as they are. For all other positions, the values ​​6b are modified to neutral values ​​in the context of the output grid 6. The modified output grid is marked with a reference mark 6*. It contains only information about the position where the object 2 is located according to the original point cloud 3*.

Claims

1. A method (100) for processing measurement data (3) from monitoring an area (1) into classification scores (2a) and / or attributes (2b) of objects (2) in the area (1) with respect to a given classification, comprising the following steps: providing (110) a point cloud (3*) of measurement data (3), the point cloud assigning a measurement value of at least one quantity to each of a plurality of points (3a-3e) in space; providing (120) an input grid (4) having a plurality of cells (4a); assigning (130) the value (4b) of the at least one measured quantity and / or the value (4b) of at least one work product derived from the at least one measured quantity to each cell (4a) of the input grid (4) based on the point cloud (3*); as well as the input grid (4) is processed (140) by the task neural network (5) into an output grid (6), the cells (6a) of which carry the classification scores (2a) and / or attributes (2b) of the objects (2) as values ​​(6b), wherein, for the input grid (4), for the output grid (6) and / or for an intermediate grid (7) that occurs during the processing of the input grid (4) into the output grid (6), Each cell (4a, 6a, 7a) is determined (131, 141) to be occupied (O) by an object (2) if at least one point (3a-3e) of the point cloud (3*) is within the cell (4a, 6a, 7a), and o if the cell (4a, 6a, 7a) has no point (3a-3e) of the point cloud (3*), it is unoccupied (U); and The values ​​(4b, 6b, 7b) assigned to unoccupied cells (4a, 6a, 7a) in the respective grid (4, 6, 7) are modified (132, 142) to neutral values ​​in the context of the respective grid (4, 6, 7).

2. The method (100) according to claim 1, wherein in the context of the respective grid (4, 6, 7), a value associated with a background without an object is selected (132a, 142a) as the neutral value.

3. The method (100) according to any one of claims 1 to 2, wherein the given classification comprises (143) in addition to the categories associated with the objects (2) a category associated with a background without objects.

4. A method (100) according to any one of claims 1 to 3, wherein the value (4b, 6b, 7b) assigned to an unoccupied cell (4a, 6a, 7a) is further modified (132b, 142b) based on the values ​​in one or more neighboring cells (4a, 6a, 7a).

5. The method (100) according to claim 4, wherein in the corresponding grid (4, 6, 7) of values ​​(4b, 6b, 7b), ● applying a softening algorithm to at least one boundary between a first region comprising occupied cells (4a, 6a, 7a) and a second region comprising unoccupied cells (4a, 6a, 7a); and / or In response to determining that a cell (4a, 6a, 7a) is occupied, adjacent cells (4a, 6a, 7a) are also considered to be occupied.

6. A method (100) according to any one of claims 1 to 5, wherein the values ​​in the intermediate grid (7) and / or the output grid (6) are selected (144) to represent pairs, which are precursors to the classification scores (2a).

7. A method (100) according to any one of claims 1 to 6, wherein during the training of the task neural network (5), the calculation of the classification scores (2a) of the elements is combined (145) with the calculation of the loss function values ​​for rating these classification scores (2a).

8. The method (100) according to any one of claims 1 to 7, wherein at most 10% of the cells in the grid (4, 6, 7) are occupied.

9. The method (100) according to any one of claims 1 to 8, wherein the task neural network (5) comprises (146) a feature extractor or object detector (5a) which produces a feature map or bounding box as an intermediate grid (7), and - At least one sorting head (5b) which further processes these intermediate grids (7) into said output grid (6).

10. The method (100) according to any one of claims 1 to 9, wherein the measurement data (3) comprises radio radar data, lidar data and / or ultrasonic data acquired using at least one sensor.

11. A method (100) according to any one of claims 1 to 10, wherein the value (4b) of the at least one measured quantity and / or the value (4b) of at least one work product derived from the at least one measured quantity is assigned (133) to each cell (4a) of the input grid using a projection neural network (8).

12. The method (100) according to any one of claims 1 to 11, further comprising: Calculating (150) an actuation signal (150a) as a function of the determined classification score (2a) and / or property (2b) of at least one object (2); as well as ● Utilizing the actuation signal (150a) to actuate (160) a vehicle (50), a traffic assistance system (51), a robot (60), a monitoring system (70), a quality inspection system (80) and / or a pick and place manufacturing system (90).

13. A computer program comprising machine-readable instructions which, when executed by one or more computers and / or computing instances, cause the one or more computers and / or computing instances to perform the method (100) of any one of claims 1 to 12.

14. A non-transitory machine-readable data carrier and / or a download product having a computer program as claimed in claim 13.

15. One or more computers and / or computing instances having a computer program as claimed in claim 13 and / or a machine-readable data carrier and / or a download product as claimed in claim 14.

Images