EP3899798A1 - Method and device for operating a machine learning model - Google Patents
Method and device for operating a machine learning modelInfo
- Publication number
- EP3899798A1 EP3899798A1 EP19809796.6A EP19809796A EP3899798A1 EP 3899798 A1 EP3899798 A1 EP 3899798A1 EP 19809796 A EP19809796 A EP 19809796A EP 3899798 A1 EP3899798 A1 EP 3899798A1
- Authority
- EP
- European Patent Office
- Prior art keywords
- data
- training
- machine learning
- learning model
- subsets
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
- 238000010801 machine learning Methods 0.000 title claims abstract description 93
- 238000000034 method Methods 0.000 title claims abstract description 56
- 238000012549 training Methods 0.000 claims abstract description 121
- 238000013528 artificial neural network Methods 0.000 claims description 51
- 230000006870 function Effects 0.000 claims description 17
- 239000011159 matrix material Substances 0.000 claims description 14
- 230000001419 dependent effect Effects 0.000 claims description 11
- 238000004590 computer program Methods 0.000 claims description 4
- 238000001514 detection method Methods 0.000 description 3
- 230000008447 perception Effects 0.000 description 3
- 238000010586 diagram Methods 0.000 description 2
- 238000013507 mapping Methods 0.000 description 2
- 230000003287 optical effect Effects 0.000 description 2
- 238000012706 support-vector machine Methods 0.000 description 2
- 241000283070 Equus zebra Species 0.000 description 1
- 230000006835 compression Effects 0.000 description 1
- 238000007906 compression Methods 0.000 description 1
- 238000013135 deep learning Methods 0.000 description 1
- 230000007613 environmental effect Effects 0.000 description 1
- 238000011156 evaluation Methods 0.000 description 1
- 238000012544 monitoring process Methods 0.000 description 1
- 210000002569 neuron Anatomy 0.000 description 1
- 238000007500 overflow downdraw method Methods 0.000 description 1
- 238000012545 processing Methods 0.000 description 1
- 238000010972 statistical evaluation Methods 0.000 description 1
- 238000007619 statistical method Methods 0.000 description 1
- 238000011144 upstream manufacturing Methods 0.000 description 1
- XLYOFNOQVPJJNP-UHFFFAOYSA-N water Substances O XLYOFNOQVPJJNP-UHFFFAOYSA-N 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F17/00—Digital computing or data processing equipment or methods, specially adapted for specific functions
- G06F17/10—Complex mathematical operations
- G06F17/16—Matrix or vector computation, e.g. matrix-matrix or matrix-vector multiplication, matrix factorization
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/40—Extraction of image or video features
- G06V10/44—Local feature extraction by analysis of parts of the pattern, e.g. by detecting edges, contours, loops, corners, strokes or intersections; Connectivity analysis, e.g. of connected components
- G06V10/443—Local feature extraction by analysis of parts of the pattern, e.g. by detecting edges, contours, loops, corners, strokes or intersections; Connectivity analysis, e.g. of connected components by matching or filtering
- G06V10/449—Biologically inspired filters, e.g. difference of Gaussians [DoG] or Gabor filters
- G06V10/451—Biologically inspired filters, e.g. difference of Gaussians [DoG] or Gabor filters with interaction between the filter responses, e.g. cortical complex cells
- G06V10/454—Integrating the filters into a hierarchical structure, e.g. convolutional neural networks [CNN]
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
- G06V10/764—Arrangements for image or video recognition or understanding using pattern recognition or machine learning using classification, e.g. of video objects
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
- G06V10/82—Arrangements for image or video recognition or understanding using pattern recognition or machine learning using neural networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V20/00—Scenes; Scene-specific elements
- G06V20/50—Context or environment of the image
- G06V20/56—Context or environment of the image exterior to a vehicle by using sensors mounted on the vehicle
- G06V20/58—Recognition of moving objects or obstacles, e.g. vehicles or pedestrians; Recognition of traffic objects, e.g. traffic signs, traffic lights or roads
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V40/00—Recognition of biometric, human-related or animal-related patterns in image or video data
- G06V40/10—Human or animal bodies, e.g. vehicle occupants or pedestrians; Body parts, e.g. hands
- G06V40/103—Static body considered as a whole, e.g. static pedestrian or occupant recognition
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N20/00—Machine learning
- G06N20/10—Machine learning using kernel methods, e.g. support vector machines [SVM]
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N7/00—Computing arrangements based on specific mathematical models
- G06N7/01—Probabilistic graphical models, e.g. probabilistic networks
Definitions
- the invention relates to a method and a device for operating a
- Machine learning and deep neural networks are also increasingly being used in vehicles, for example in infotainment systems,
- Functions of deep learning are used, in which higher-value data, for example in the form of an environment model, object detection, regulation or a driver model, are generated from sensor data recorded using a sensor (e.g. environment sensors, interior monitoring, sensors in or on the vehicle, etc.).
- a sensor e.g. environment sensors, interior monitoring, sensors in or on the vehicle, etc.
- Machine learning models and especially deep neural networks are in use, i.e. during an inference phase when running in the vehicle, very computation-intensive.
- the invention is based on the object of providing a method and a device for operating a machine learning model in which the machine learning model requires less effort, in particular with regard to one required during an inference phase
- Computing power can be operated.
- a method for operating a machine learning model comprising the following steps during a training phase:
- Machine learning model on each stack element of the data stack, the application taking place simultaneously, and deriving an inference result
- a device for operating a machine learning model comprising a computing device with a memory in which the machine learning model is formed, the computing device being set up to carry out the following steps during a training phase:
- the training data record comprising data record elements which are generated on the basis of the selected subsets
- Machine learning model on each stack element of the data stack, the application taking place simultaneously, and deriving an inference result
- the subsets can be image sections of an image captured by the camera.
- the image sections have fewer pixels than the original image, so that less input data from both during training and when used
- Machine learning model must be processed. Therefore, subsets are selected from marked (labeled) multi-dimensional training data. From the selected
- a training data record is generated for subsets, wherein individual data record elements of the training data record are each generated on the basis of one of the selected subsets. If, for example, five pedestrians are depicted in an image of the camera in the aforementioned example, they can each be cut out of the image as an image section or subset and can be included in the training data record as a data record element.
- the machine learning model is then trained using the training data set provided. Since the subsets each are less on their own
- Input dimensions are made smaller, that is, it is compressed. This can reduce the computing power required to use the machine learning model during an inference phase.
- sensor data are received by at least one sensor. Subsets are also selected from these sensor data.
- these can again be image sections of an image captured by a camera.
- a data stack is generated, the data stack each comprising the selected subsets, for example the image sections already described, as stack elements.
- Training phase of the method trained machine learning model is then applied to each stack element of the data stack. This takes place in parallel or simultaneously, so that the trained machine learning model is applied to a stack element of the data stack.
- the trained machine learning model is in particular instanced or generated several times. By applying them in parallel, inferred results are available for everyone
- Inference phase even if the machine learning model is applied to all stacking elements at the same time. As a result, a required computing power can be reduced both in the training phase and in the inference phase.
- the method and the device can be used in particular in the context of an environment detection on the basis of acquired sensor data or image data of the environment.
- the method and the device can be used in a vehicle, for example to support an assistance system in automated driving.
- a vehicle is in particular a motor vehicle, in particular a partially automated or automated motor vehicle.
- the vehicle can also be another land, air or water vehicle.
- the method is divided into a training phase in which the machine learning model is trained and an inference phase in which a trained machine learning model is applied to sensor data.
- the training phase and the inference phase are not carried out using the same device.
- a machine learning model can be trained on a device by a manufacturer of a vehicle.
- the machine learning model that has already been trained is then installed in a further device for providing a perception function in an assistance system of the vehicle.
- the trained machine learning model is used there, for example, to recognize objects in sensed sensor data in an environment of the vehicle. For this purpose, only the inference phase is carried out in the device.
- Process can be carried out on a single device, for example in one
- the method is carried out in particular by means of a computing device with a memory.
- the machine learning model is formed in particular in the memory and the computing device is set up to adapt the weights or parameters etc. of the machine learning model as a function of an inferred result and at least one target function during training.
- the input data of the machine learning model i.e. the training data and the
- Sensor data can be images captured by a camera, for example, individual video images. However, input data can also be recorded using other sensors, for example using a radar, light detection and ranging (lidar) or ultrasonic sensor.
- sensors for example using a radar, light detection and ranging (lidar) or ultrasonic sensor.
- the machine learning model processes the input data, i.e. training data and
- Sensor data in particular image data, in a lower resolution than that of one Sensor are provided.
- the resolution is reduced.
- a variable resolution of the input data or a fixed resolution can be provided.
- Neural networks is a resolution of the input data with so-called full
- Convolutional networks for example, are variable. However, other topologies of deep neural networks require a fixed resolution of the input data. The reduced input data represent the partial quantities in the process.
- the training data or sensor data provided by the at least one sensor can be reduced in various ways to a resolution processed in the input data by the machine learning model.
- downsampling can reduce both triggering, e.g. a number of pixels, as well as a number of markings (labels) can be achieved in the training data.
- segments with a corresponding resolution can be cut out of the training data or the sensor data. If it is, for example, image data, segments with the appropriate resolution, size or number of pixels can be cut out of the image data. It can be provided that the segments are cut out of an image center and / or along a horizon.
- the individual inferred results of the stacking elements that are output as stacks are then combined or combined into a single result.
- the subsets represent image sections of an image captured by means of a camera
- at least the part of the image which was taken into account by selecting the respective subsets can then be reassembled and provided as an overall result, that is to say in particular as a single model prediction become.
- the combination can be carried out, for example, using appropriate masks, so that the subsets can be inserted at the original position of the sensor data or the image. It can be provided here that the subsets are shifted with respect to one position and / or with respect to one
- Resolution are scaled to be moved back to an original position and / or to be brought to an original resolution.
- machine learning model When applying the machine learning model to the data stack, it may happen that individual stack elements overlap with respect to the subsets contained therein, that is to say that the respective subsets have a common intersection.
- image data used as subsets for example, image sections in an image can overlap with one another.
- the inferred results delivered for the individual subsets can be combined to form a result for overlapping areas, in particular merged with one another. This can be done in different ways. Provision can be made to offset the overlapping regions with one another, for example by forming an average, a weighted average or locally weighted averages. It can also be provided
- a suitable mask for summarizing, for example the smallest, largest or an object-specific mask.
- a decision can also be made on the basis of a confidence measure determined for the respective subset involved or a prediction result. E.g. the most confident or the most common prediction can be used.
- a distance to a mask or image center and / or a mask or image edge can also be used. Fusion methods learned by machine learning can also be used, for example methods that merge the results based on a context, semantics or training experience.
- the machine learning model can be any suitable model created by means of machine learning, which can be trained in a training phase on the basis of marked training data and can then be applied to sensor data in an inference phase.
- suitable machine learning models are support vector machines, neural networks or probabilistic models, for example based on Bayesian networks.
- the machine learning model is a deep neural network.
- the selection of the subsets takes place during the training phase and / or during the inference phase on the basis of a relevance of the respective subsets.
- the training data and the sensor data For example, around image data from a camera, only relevant areas can be taken into account in this image data. In this way, attention can be focused in the input data processed by the machine learning model.
- a relevance arises, for example, by considering criteria such as security (eg weak road users in the training data and / or in the sensor data) or particularly high-contrast or low-contrast image areas.
- a list with relevant subsets is created for training data and sensor data and the selection is then made on the basis of the created list, for example by only selecting the most relevant 5, 10, 20, ... subsets .
- the partial quantities are selected from the training data and / or the received sensor data additionally or alternatively on the basis of a situation-dependent context. If, for example, the training data and the sensor data are image data from a camera, it can be provided that, depending on a situation-dependent context, different ones
- Image sections can be selected as subsets.
- the selected subsets can be selected differently when driving on a motorway than when driving around town.
- lanes running to the left and right of a vehicle are of increased interest since they provide potential alternatives for the vehicle.
- a bicycle lane for example, may be of increased interest when driving around the city, since there may be a weaker road user on it.
- the subsets are accordingly selected depending on the situation. In this way, situations that are identified as critical can also be better considered. For example a zebra crossing or a child playing on a road ahead can be taken into account better by selecting these image sections as subsets.
- a planned route is evaluated as a context or is used to determine a situation-dependent context for at least one current point in time.
- the route can be queried, for example, by a navigation device of a vehicle.
- the respective subsets can then be selected.
- other information can also be taken into account, such as a road condition, current or future weather, weather conditions, traffic jam reports, etc.
- the selection of the subsets or the generation of the data stack is made dependent on results of the inference phase of a method carried out at an earlier point in time.
- expected or predicted shifts of relevant subsets in the sensor data can be taken into account. For example, movements and / or perspective
- At least one confidence measure is determined and / or output for the machine learning model, the at least one confidence measure being determined and / or output separately for at least two selected subsets and / or stacking elements.
- a confidence measure can be specified separately for each of the subsets or each of the stacking elements.
- the procedure must be recorded and / or evaluated more precisely.
- the confidence measure is, in particular, a confidence value with regard to the respective inference result, for example an object class or an object size, etc.
- a confidence value is an indication of the probability with which the trained machine learning model can determine an object under consideration or the corresponding size.
- the confidence measure thus defines the correctness of a perception or an inferred result of the trained machine learning model. If, for example, an object class is pedestrians, the confidence value when recognizing a pedestrian indicates that the trained machine learning model can recognize the object "pedestrian" with a probability of, for example, 99%. If the machine learning model is a deep neural network, it can
- Confidence measure can be determined, for example, by statistical evaluation of inference results when the trained deep neural network is used repeatedly on the same or similar input data. Will with the deep neural network
- a confidence for the inferred result can be achieved by multi-inference, i.e. by repeatedly applying the deep neural network to the same input data , determine.
- the multiple inferred results are evaluated using statistical methods and derived from them
- the data record elements in the training data record are combined into a matrix, the computing operations required for training being carried out on the matrix and / or for using the trained machine learning model, the stack elements in the data stack are combined into a matrix , the necessary for inferring
- Arithmetic operations are performed on the matrix.
- matrix is intended to refer in particular to a hypermatrix, that is to say a matrix with more than two indices. Is the training data and sensor data, for example?
- the selected subsets are also two-dimensional.
- the two-dimensional data for example image data from an image of a camera.
- Subsets are then combined into a three-dimensional matrix.
- the training and the application are then carried out on the three-dimensional matrix.
- a data resolution at the input of the machine learning model and / or a number of stack elements of the data stack is determined as a function of a computing power available when the trained machine learning model is used and / or a maximum possible latency.
- a machine learning model can be tailored for an application scenario. For example, a maximum latency of 30 milliseconds is provided for a perception function in a vehicle, since then on the basis of a
- the machine learning model is designed in such a way that the application falls below the 30 milliseconds during the inference phase.
- a data resolution for example a number of pixels in image data, and / or a number of stack elements of the data stack can be changed.
- a data resolution for example a number of pixels in image data, and / or a number of stack elements of the data stack can be changed.
- a compromise between computing power or computing time and a quantity of processed input data is sought.
- One embodiment provides that when training the machine learning model, a target function for at least one of the subsets is selected or specified separately.
- the machine learning model can be trained specifically for certain properties of the subset.
- loess or cost functions dependent on the image section can be used, for example. The advantage is that this can speed up the training phase.
- a vehicle comprising at least one device according to one of the described embodiments.
- a computer program with program code means is created in order to carry out all steps of the method in one of the described embodiments when the program is executed on a computer.
- a computer program product is created with program code means which are stored on a computer-readable data carrier in order to carry out the method according to one of the described embodiments when the program product is executed on a computer.
- Fig. 1 is a schematic representation of an embodiment of the device for
- Fig. 2 is a schematic representation of an embodiment of the device for
- Fig. 3 is a schematic flow diagram of an embodiment of the method for
- FIG. 1 shows a schematic representation of an embodiment of the device 1 for operating a machine learning model 6 during a training phase.
- machine learning model 6 is a deep neural network 4.
- the device 1 comprises a computing device 2 with a memory 3.
- the deep neural network 4 is formed in the memory 3, i.e. its structural characteristics and associated weights are stored there.
- Computing device 2 selects multi-dimensional training data 10. Subsequently, computing device 2 selects subsets 11 from training data 10. For this purpose, the computing device 2 comprises a selection device 5.
- the selection device 5 selects the subsets 11 based on a relevance 14 of the respective subsets 11. Requirements for the
- Relevance 14 can be provided to the selection device 5 from the outside, for example.
- the selection device 5 additionally or alternatively selects the subsets 11 based on a situation-dependent context 15.
- a current context 15 can be provided to the selection device 5 from the outside, for example.
- the selection device 5 then provides a training data record 12, the training data record 12 being composed of data record elements 13 which were each generated by the selection device 5 on the basis of the selected subsets 11.
- the deep neural network 4 is then trained using the training data set 12 compiled in this way.
- the training is carried out by the computing device 2. It can be provided that at least one confidence measure is determined and / or output for the deep neural network 4 after the training.
- the confidence measure specifies in particular the probability with which the deep neural network 4 can correctly recognize a specific result after training, for example that the neural network 4 can correctly recognize the object “pedestrian” after training with a probability of 98%.
- FIG. 2 shows a schematic representation of an embodiment of the device 1 for operating a trained machine learning model 6a during an inference phase, the trained machine learning model 6a being a trained deep neural network 4a.
- the device 1 can be designed to carry out the inference phase or to use the trained deep neural network 4a, in particular in a vehicle 50.
- the device 1 comprises a computing device 2 with a memory 3.
- the trained deep neural network 4a is formed in the memory 3, i.e. its structural characteristics and the weightings determined during the training phase (see FIGS. 1 and 3) are stored there.
- Computing device 2 multidimensional sensor data 20 of a sensor 52.
- the sensor 52 is, for example, a top view camera of the vehicle 50, which provides image data in the form of captured images of an environment of the vehicle 50 as sensor data 20.
- the computing device 2 then selects from the received sensor data 20
- the computing device 2 comprises a selection device 5.
- the selection device 5 selects the subsets 21 based on a relevance 14 of the respective subsets 21. Requirements for the
- Relevance 14 can be provided to the selection device 5 from the outside, for example.
- the selection device 5 additionally or alternatively selects the subsets 21 on the basis of a situation-dependent context 15.
- a current context 15 can be provided to the selection device 5 from the outside, for example.
- the specifications for relevance 14 and / or for the situation-dependent context 15 during the inference phase correspond to the specifications during the training phase.
- the selection device 5 generates a data stack 22 from the selected subsets 21, the data stack 22 each comprising the selected subsets 21 as stack elements 23.
- the data stack 22 is fed to the trained deep neural network 4a.
- the trained deep neural network 4a is then applied to the data stack 22.
- the computing device 2 uses several instances of the trained depth
- Neural network 4a generated simultaneously, a number of a number of
- Computing device 2 simultaneously assigns one of the stack elements 23 to the individual instances of the trained deep neural network 4a.
- the instances of the trained deep neural network 4a are the instances of the trained deep neural network 4a.
- Neural networks each deliver an inferred result.
- the infered results are also provided as a stack or in a summarized form as an inference result 24.
- the inference result 24 is then output, for example as a digital one
- the output takes place, for example, using a device provided for this purpose
- the output inference result 24 can then be processed further, for example by an assistance system 52 of the vehicle 50 as part of an environmental interpretation or for planning a trajectory of the vehicle 50.
- At least one confidence measure 16 is determined and / or output for the trained deep neural network 4a. It is provided here that the at least one confidence measure 16 is determined and / or output separately for at least two stacking elements 23.
- a confidence value 16 can be specified for each result inferred for a stack element 23, that is to say in particular an indication of the probability with which the trained deep neural network 4a correctly recognizes the inferred result in each case, that is to say a measure of the correctness or
- FIG. 3 shows a schematic flow diagram of an embodiment of the method for operating a machine learning model.
- the machine learning model is a deep neural network. There are two parts to the process
- Method steps 201 to 204 are carried out. Marked multi-dimensional training data are received in a method step 201. This can be, for example, images of an environment captured by a camera, for example an environment of a vehicle. The illustrations are marked here
- the images are assigned information about the objects present in the images and / or their size, properties, etc.
- subsets are selected from the training data.
- these are image sections from the illustrations. If, for example, an image of a street scene is involved, image sections with pedestrians can be selected as subsets.
- the subsets or image sections can then be processed further, in particular a resolution or number of pixels can be adapted to an input dimension of the deep neural network.
- the selection of the subsets takes place during the training phase on the basis of a relevance of the respective subsets.
- the subsets or image sections can be selected depending on a security-relevant property of the depicted object.
- weak road users such as Pedestrians are rated as highly relevant and should therefore be considered when selecting the subsets or image sections.
- Properties of an image itself can also be used to derive relevance. For example, low-contrast and therefore difficult to see image sections can be given high relevance so that they are selected as subsets.
- a data resolution at the input of the deep neural network is determined as a function of a computing power available when the trained deep neural network is used and / or a maximum possible latency.
- the available computing power and / or the maximum possible latency are predefined, for example, by hardware of a vehicle.
- the data resolution is then selected in such a way that the available computing power and / or the maximum possible latency is never fully utilized.
- a training data record is provided in a method step 203.
- Training data record comprises data record elements, each of which is selected from the selected
- Partial quantities are generated. Since a selection has been made, it is no longer necessary to use a complete (especially high-resolution) image for training the deep neural network. Training takes place only with the help of the individual
- the deep neural network is trained using the training data record.
- Known methods of machine learning can be used here.
- a target function can be used specifically for certain subsets or image sections. This can accelerate a training phase.
- the inference phase 300 can be carried out independently of the training phase 200. For example, it can be provided that the inference phase 300 in a vehicle for Providing an assistant function is carried out. A deep neural network trained according to the training phase 200 is used for this. The trained deep neural network is stored, for example, by a manufacturer of the vehicle in the memory
- the device as described in FIG. 2 is deposited and can then be used in a subsequent inference phase 300 when the vehicle is delivered to a customer.
- the inference phase 300 comprises the method steps 301 to 305.
- sensor data are received from at least one sensor.
- the sensor data are, for example, an image of an environment of the vehicle captured by means of a camera.
- subsets are selected from the received sensor data.
- the subsets are image sections of the illustration.
- the subsets or image sections can be selected depending on a security-relevant property of an object depicted therein. For example, weak ones
- the partial quantities are selected from the sensor data on the basis of a situation-dependent context. For example, depending on the current situation, other subsets or image sections can be taken into account when selecting. Areas on lanes of the freeway are particularly relevant on a freeway; traffic lights and pedestrians are generally not to be found there. In city traffic, however, traffic lights, pedestrians and other objects, especially weaker road users, should be given priority when selecting.
- Inference phase serve as the starting point for the selection. If an inference result shows, for example, that a pedestrian can be seen in an image, the associated image section can also be selected in an image captured at a later time. It can be provided here that the selection of the subsets or
- Image sections based on salience processes human salience or machine salience, e.g. optical flow, trained salience processes, heat mapping processes in a previous time step, etc.).
- a data stack is generated.
- the data stack comprises stack elements which are each generated from the selected subsets.
- a number of stack elements of the data stack is determined as a function of a computing power available when the trained deep neural network is used and / or a maximum possible latency.
- the available computing power and / or the maximum possible latency are specified, for example, by a vehicle's hardware.
- the number of stack elements is then selected such that the available computing power and / or the maximum possible latency is never fully utilized.
- the trained deep neural network is applied to the data stack.
- an instance of the trained deep neural network is generated for each stack element and then one of the stack elements is fed to each instance.
- the individual instances then deliver an inferred result for each stack element, that is to say for each subset or each image section.
- the inferred results become one
- the combination can also include, for example, a mask-based generation of a single image from the subsets or the individual inferred results, so that a single image with the associated inferred results can be provided. For example, an image with objects classified therein and associated object positions can be provided in this way.
- the inference result is output in a method step 305, for example as a digital data packet.
- the inference result or the digital data packet can then be made available to an assistance system of a vehicle.
- the method is then ended 400.
- the inference phase of the method is then repeated, sensor data recorded at a later point in time being evaluated.
- the machine learning model 6, 6a is a deep neural network 4, 4a.
- the machine learning model 6 can also be designed differently, for example in the form of a support vector machine or in the form of a probabilistic model.
- the training phase and the inference phase are then carried out analogously for these machine learning models 6, 6a.
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Evolutionary Computation (AREA)
- Multimedia (AREA)
- Artificial Intelligence (AREA)
- Health & Medical Sciences (AREA)
- General Health & Medical Sciences (AREA)
- Software Systems (AREA)
- Computing Systems (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Data Mining & Analysis (AREA)
- Mathematical Physics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Databases & Information Systems (AREA)
- General Engineering & Computer Science (AREA)
- Molecular Biology (AREA)
- Biomedical Technology (AREA)
- Medical Informatics (AREA)
- Computational Mathematics (AREA)
- Pure & Applied Mathematics (AREA)
- Mathematical Optimization (AREA)
- Mathematical Analysis (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Human Computer Interaction (AREA)
- Biodiversity & Conservation Biology (AREA)
- Algebra (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Evolutionary Biology (AREA)
- Bioinformatics & Computational Biology (AREA)
- Image Analysis (AREA)
- Traffic Control Systems (AREA)
Abstract
Description
Claims
Applications Claiming Priority (2)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
DE102018222202.9A DE102018222202A1 (en) | 2018-12-18 | 2018-12-18 | Method and device for operating a machine learning model |
PCT/EP2019/082486 WO2020126339A1 (en) | 2018-12-18 | 2019-11-25 | Method and device for operating a machine learning model |
Publications (1)
Publication Number | Publication Date |
---|---|
EP3899798A1 true EP3899798A1 (en) | 2021-10-27 |
Family
ID=68699438
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
EP19809796.6A Pending EP3899798A1 (en) | 2018-12-18 | 2019-11-25 | Method and device for operating a machine learning model |
Country Status (4)
Country | Link |
---|---|
US (1) | US20220044118A1 (en) |
EP (1) | EP3899798A1 (en) |
DE (1) | DE102018222202A1 (en) |
WO (1) | WO2020126339A1 (en) |
Families Citing this family (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
DE102020209900A1 (en) | 2020-08-05 | 2022-02-10 | Robert Bosch Gesellschaft mit beschränkter Haftung | Method and analysis device for processing environmental sensor data |
US11943244B2 (en) | 2021-06-22 | 2024-03-26 | International Business Machines Corporation | Anomaly detection over high-dimensional space |
DE102022209460A1 (en) | 2022-09-09 | 2024-03-14 | Volkswagen Aktiengesellschaft | Method for controlling an artificial neural network when operating a vehicle |
Family Cites Families (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US10373019B2 (en) * | 2016-01-13 | 2019-08-06 | Ford Global Technologies, Llc | Low- and high-fidelity classifiers applied to road-scene images |
US10592805B2 (en) * | 2016-08-26 | 2020-03-17 | Ford Global Technologies, Llc | Physics modeling for radar and ultrasonic sensors |
US10360494B2 (en) * | 2016-11-30 | 2019-07-23 | Altumview Systems Inc. | Convolutional neural network (CNN) system based on resolution-limited small-scale CNN modules |
-
2018
- 2018-12-18 DE DE102018222202.9A patent/DE102018222202A1/en active Pending
-
2019
- 2019-11-25 WO PCT/EP2019/082486 patent/WO2020126339A1/en unknown
- 2019-11-25 US US17/414,675 patent/US20220044118A1/en active Pending
- 2019-11-25 EP EP19809796.6A patent/EP3899798A1/en active Pending
Also Published As
Publication number | Publication date |
---|---|
US20220044118A1 (en) | 2022-02-10 |
DE102018222202A1 (en) | 2020-06-18 |
WO2020126339A1 (en) | 2020-06-25 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
DE112017006530B4 (en) | FEEDBACK FOR AN AUTONOMOUS VEHICLE | |
EP3765927B1 (en) | Method for generating a training data record for training an artificial intelligence module for a control device of a vehicle | |
EP3899798A1 (en) | Method and device for operating a machine learning model | |
DE102018132520A1 (en) | Method and system for controlling a motor vehicle | |
DE102016215314A1 (en) | Driver assistance system, means of transportation and method for predicting a traffic situation | |
EP3789926A1 (en) | Method for detecting an adversarial fault in the input data of a neural network | |
WO2018215242A2 (en) | Method for determining a driving instruction | |
DE102018203745A1 (en) | Method, a device and a corresponding computer program for varying or suppressing at least one driver information | |
EP4078238A1 (en) | Method and device for making sensor data more robust against adverse disruptions | |
EP3935463B1 (en) | Method and device for operating an automated vehicle | |
EP4350660A1 (en) | System and method for predicting a future position of a traffic participant | |
DE102019204187A1 (en) | Classification and temporal recognition of tactical driving maneuvers by road users | |
DE102018216719A1 (en) | Keyframe-based autonomous vehicle operation | |
AT524932B1 (en) | Method and system for testing a driver assistance system for a vehicle | |
DE102018203743A1 (en) | A method, a device, a mobile user device, and a corresponding computer program for varying or suppressing at least one driver information | |
WO2021078512A1 (en) | Method for making a neural network more robust against adversarial disruptions | |
DE102013219414A1 (en) | Method and device for providing predictive vehicle movement information for a driver assistance system of a vehicle | |
DE102020129451A1 (en) | Method for predicting driving interventions, method for training an algorithm and motor vehicle | |
EP4078237A1 (en) | Method and apparatus for recognising removal of a sensor data domain from a reference data domain | |
DE102019104973A1 (en) | Method and control device for a system for controlling a motor vehicle | |
DE102020201016A1 (en) | Method and device for providing at least one trajectory for an automatically driven vehicle | |
DE102018203742A1 (en) | A method, a device, a mobile user device, and a corresponding computer program for varying or suppressing at least one driver information | |
DE102019118607A1 (en) | ANOMALY DETECTOR FOR VEHICLE CONTROL SIGNALS | |
DE102006004361A1 (en) | Method for determination of duration of vehicle presence at specific location, comprises use of reference times and -paths | |
DE102019213222B4 (en) | Method for predicting a future driving situation of a foreign object, device, vehicle participating in road traffic |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
STAA | Information on the status of an ep patent application or granted ep patent |
Free format text: STATUS: UNKNOWN |
|
STAA | Information on the status of an ep patent application or granted ep patent |
Free format text: STATUS: THE INTERNATIONAL PUBLICATION HAS BEEN MADE |
|
PUAI | Public reference made under article 153(3) epc to a published international application that has entered the european phase |
Free format text: ORIGINAL CODE: 0009012 |
|
STAA | Information on the status of an ep patent application or granted ep patent |
Free format text: STATUS: REQUEST FOR EXAMINATION WAS MADE |
|
17P | Request for examination filed |
Effective date: 20210719 |
|
AK | Designated contracting states |
Kind code of ref document: A1 Designated state(s): AL AT BE BG CH CY CZ DE DK EE ES FI FR GB GR HR HU IE IS IT LI LT LU LV MC MK MT NL NO PL PT RO RS SE SI SK SM TR |
|
DAV | Request for validation of the european patent (deleted) | ||
DAX | Request for extension of the european patent (deleted) | ||
STAA | Information on the status of an ep patent application or granted ep patent |
Free format text: STATUS: EXAMINATION IS IN PROGRESS |
|
17Q | First examination report despatched |
Effective date: 20240222 |