DE102022107820A1 - Method and processor circuit for checking the plausibility of a detection result of object recognition in an artificial neural network and motor vehicle - Google Patents

Abstract

The invention relates to a method for checking the plausibility of a detection result of an object recognition in an artificial neural network (25), wherein input data (24) relating to an object (16) is received by the network (25) and as a function of activation values of the network (25 ) Detection data (18) are generated, which indicate which object class (20) from several predetermined object classes (20) detectable by the network (25) was detected for the object (16). The invention provides that, in addition to the detection data (18), a feature vector (43) is formed from a predetermined set (38) of activation values and a statistical frequency model (45) assigned to the detected object class (20) is selected and by means of the frequency model ( 45) a class probability information (47) is determined, which indicates how likely the feature vector (43) is for the detected object class (20).

Description

The invention relates to a method and a processor circuit for checking the plausibility of a detection result of an object recognition that is carried out by an artificial neural network. The network can be trained to recognize an object of a predetermined object class from input data (e.g. an image section) and to signal the recognized or detected object class. However, such a detection result can be subject to uncertainty that cannot be attributed to the detection result itself. Therefore, it should also be checked whether the detection result can be plausible at all given the input data. The invention also relates to a motor vehicle in which an automated driving function is supported by object recognition using the processor circuit.

In a motor vehicle, object recognition can be operated to recognize or detect objects in the environment, which can be based on an artificial neural network, in particular a multi-layer, so-called deep neural network (DNN - Deep Neural Network). In preprocessing for the network, a camera image can be used to identify which image areas show a possible object and such an image area can then be marked, for example, by a bounding box. The image data of such a bounding box or such an image area can then represent the input data for the artificial neural network, which carries out object recognition on the input data and, as the recognition result, indicates an object class that identifies the object, for example as an object class “vehicle” or “pedestrian”. or “road sign”, to name just examples of object classes.

By means of such a detection result, an automated driving system or an automated driving function, for example an autopilot or a driver assistance system, can be supported or such an automated driving function can plan a driving trajectory and/or a braking maneuver based on the detection result. In general, semantic segmentation or 2D/3D object detection or image classification can be carried out using an artificial neural network through object recognition.

An artificial neural network can be trained to distinguish between several different classes of objects. To train or teach an artificial neural network, annotated or labeled training data can be used, which additionally indicates the so-called ground truth, i.e. the true or actual object class, in addition to input data, as it can be specified by an operator, for example. In a known manner, a network can be trained using a training algorithm, for example a backpropagation algorithm.

In principle, an artificial neural network can always give a result for all object classes for given input data and the most likely object class is then output as the detection result, for example using a softmax function.

A disadvantage here is that the detection result itself cannot be seen as to how “close” the artificial neural network’s decision is that the object class actually exists. For example, if several object classes (e.g. C1, C2, C3) are almost equally likely (e.g.: C1: 32%, C2: 32%, C3: 36%) and the probability of detection for one object class is only marginal or minimal (in the example C3 with only 4 percentage points), this object class (C3) can still be specified as the detection result. How reliable this detection result is or how “uncertainty” the artificial neural network is with this detection result cannot be determined.

Calculating an “uncertainty” of an artificial neural network given a given detection result can be computationally very complex.

In addition, a detection result can even be wrong because, given similar detection probabilities, even a small amount of noise or a small amount of scatter can have a decisive influence.

The use of an artificial neural network to support driving of a motor vehicle is, for example, from DE 197 20 764 A1 known.

From the DE 10 2006 003 611 A1 a method is known for checking data sets for their consistency using a statistical analysis, i.e. a single data set that represents an outlier in terms of the statistical distribution can be identified using the method.

From the DE 10 2018 211 099 A1 It is known that an artificial neural network can be used to recognize additional parameters relating to a measurement situation how reliable a measurement of a measured value can actually be.

From the DE 10 2019 116 801 A1 Another artificial neural network is known, by means of which the driving operation of a motor vehicle can be controlled depending on measurement data.

The invention is based on the object of checking a detection result of an artificial neural network for plausibility or reliability.

The task is solved by the subject matter of the independent patent claims. Advantageous further developments or further developments of the invention are described in the dependent claims, the following description and the figures.

As a solution, the invention includes a method for checking a plausibility or reliability of a detection result that an artificial neural network has output from object recognition. The method assumes that input data relating to an object is received by the network. This is an object that is detected by sensors in an environment, for example by means of an environmental camera of a motor vehicle. The input data can be, for example, the image data from the area of a bounding box of the image that was received from the surrounding camera. The bounding box itself can be determined, for example, by an upstream image analysis circuit in a manner known per se.

In the network, activation values of artificial neurons in the network are calculated or determined in several layers of the network as a function of the input data. This is known from the prior art. As a function of these activation values, recognition data or detection data are then generated as a detection result at an output layer of the network. The detection data indicates which object class was detected for the object. The object class is one of several predetermined object classes that can be detected by the network. The detection data can, for example, specify a label or an annotation text that can describe the object class (e.g. “car” or “traffic sign” or “road marking”, to name just examples). The detection result, i.e. the detected object class, can be the dominant result to which the greatest probability of detection has been assigned. In other words, a (smaller) detection probability can also have been recognized for at least one other object class to be present. Such a result can be concealed or hidden, for example, by a Softmax layer. In other words, the set of detection values or detection probabilities can have different values for several object classes. The smaller the entropy of these multiple recognition values, the more reliable the recognition result is, because this indicates that a certain object class is dominant in the recognition (e.g. in the example above, C1: 2%, C2: 2%, C3: 96% a lower entropy than the first example above).

In order to check the detection result, i.e. to verify the detected object class, a feature vector is formed by a processor circuit in addition to the actual detection data. The feature vector is formed from a predetermined set of activation values of at least one of the layers of the network. It is checked whether the feature vector is “typical” or “probable” if the detected object class is actually present, i.e. if the input data actually represents an object of this object class. For this purpose, a statistical frequency model assigned to the detected object class is selected as a function of the detection data. The frequency model is a statistical model for feature vectors that arise when an object of the detected object class is actually depicted in the input data. Using the frequency model, a class probability information is then determined, which indicates how likely the feature vector is for the object class.

By means of the processor circuit, a false detection can then be recognized (detected object class does not match the actual object class of the object) and, if necessary, reacted to by mapping the class probability information using a predetermined mapping function onto a control signal, which controls a predetermined security measure to compensate for a false detection. The class probability information can be a binary information (detection result correct or incorrect) or a probability value (feature vector corresponds to X% of the object class). This class probability information can then be represented in a signal value of the control signal. The control signal can, for example, specify whether the security measure should be carried out (false detection detected) or whether no security measure is necessary (detection result recognized as correct or correct).

The invention has the advantage that, independently of the object detection by the artificial neural network, it is also checked whether the object belongs to the detected object class using a frequency model that is independent of it. If a contradiction is detected (detected object class only gives a low class probability information for the feature vector), a security measure is triggered so that the false detection is no longer used or the effects of the false detection are alleviated or prevented. The frequency model mentioned can be generated using the training data mentioned at the beginning, for example using a Gaussian kernel model.

The invention also includes further developments or further developments that result in additional advantages.

A further development includes that the set of activation values is taken from a layer or several layers, each of which is an output of a feature extractor of the network, which is followed by a detector of the network. Another name for “feature extractor” is “encoder”, which can be formed from several layers of artificial neurons and can be designed as a so-called convolutional network CNN (Convolutional Neural Network). Using the feature extractor, independently of any actual recognition or classification, the features contained therein are first extracted from the input data, for example edges and/or textures, which form the basis for the actual recognition or detection by the downstream detector. The detector itself can, in a manner known per se, be an arrangement of several layers of artificial neurons in, for example, a so-called “Fully-Connected (FC) Layers” network. The transition between feature extractor and detector is also called a bottleneck. This arrangement of feature extractor and detector is known from the prior art. It was recognized that this transition is particularly suitable for using the activation values (output values of their activation functions) generated by the artificial neurons arranged there to form the feature vector. The feature vector is therefore formed, in particular, independently of the actual detector, so that a check of the detection result is independent of a false detection in the detector itself. The “activation values” described here represent the output values of the individual artificial neurons, which are known from the prior art, as they can be generated or output by the non-linear activation function of an artificial neuron.

A further development includes that a dimension of the feature vector is smaller than a number of activation values in the said set of activation values and for this purpose the set of activation values is mapped onto the feature vector using a linear model for multivariable statistics or is converted into the feature vector. This results in the advantage that a large number of activation values, for example the activation values of an entire layer or all recognized features, can be taken into account in the feature vector and yet the frequency model for testing the feature vector (e.g. a Gaussian distribution) is based on a smaller number of variables can be generated. The frequency model can be formed, for example, using the training data, which is also used to train the artificial neural network, as a statistical distribution of the vector values of the feature vector that result for the training data. A linear model for multivariable statistics represents a dimensionality reduction model in which the dimension of the feature vector is less than the dimension of a vector containing the set of selected activation values. An example of a linear model for multivariable statistics to generate the feature vector is a cluster analysis or factor analysis.

Principal Compound Analysis (PCA) has proven to be a particularly preferred model. A further development includes using principal component analysis, PCA, as the linear model of multivariable statistics. This has made it possible to map a large number of activation values, for example more than 50 or more than 100, onto a relatively small number of vector values for the feature vector, in particular less than 10. You still get a reliable or meaningful frequency model for checking the detected object class.

A further development includes that the feature vector has dimensionality 3. With a dimensionality of 3, it is possible to also represent the frequency model and the feature vector graphically, allowing visual inspection by an operator. Thus, for example, during development or programming of the processor circuit, a developer can visually check the reliability or functionality of the method.

A further development includes that the respective frequency model includes a kernel-based distribution function of a frequency distribution, in particular a Gaussian kernel-based distribution function, and a probability value or probability density value is determined as class probability information. The frequency model can therefore advantageously be modeled as a statistical distribution of the vector values of the feature vector for objects of the respective object class, for example through relative frequencies. Since such a frequency model usually does not have a symmetrical arrangement or shape, a single (single) Gaussian distribution is usually too coarse. Here it was recognized that a Gaussian kernel distribution function is more suitable for appropriately approximating asymmetric frequency distributions. Using a distribution function, a probability value can be specified as class probability information, which can be used to quantify how clear the distinction between the detected object class and the next similar object class is.

Additionally or alternatively, the respective frequency model includes a support vector machine, SVM, for a classification analysis of the feature vector and a binary membership statement (belongs to the object class or not) is determined as the class probability information. The use of a support vector machine has the advantage that it is used to decide whether a feature vector is part of the frequency model or not, i.e. there is a binary class probability information that can be used directly to generate the control signal (activate safeguarding measures or not). The mapping function in this case can be the identity function (output = input).

Once the class probability information is determined, a decision must still be made based on this as to whether the hedging measures should be triggered. This is done using the mapping function mentioned. A further development includes using the mapping function to check whether the feature vector confirms the object with a predetermined minimum probability according to the class probability information. In the event that the class probability information represents a probability value, it can be controlled for which feature vectors a false detection or false detection should be detected. This can be done by setting the minimum probability value, i.e. setting the limit. This allows the detection behavior or checking to be calibrated or adjusted.

A further development includes that the security measure includes increasing a spatial resolution and/or a time resolution for at least one environmental sensor that generates the input data (if a false detection is detected, i.e. the class probability information contradicts the detected object class). In other words, the detection can continue to be operated, but a measurement quality or a measurement scope is automatically increased for the input data. This advantageously results in automated scaling or adjustment when operating the environmental sensor depending on the reliability or plausibility of the detection results. Additionally or alternatively, the security measure includes connecting predetermined computing resources to generate the input data. In this way, the quality of the generation of the input data can be improved automatically, for example by increasing computing resources for calculating or determining bounding boxes and thus, for example, increasing or improving the accuracy in the delimitation of objects that are imaged in a camera image. This results in automated scaling of computing resources when generating the input data. Additionally or alternatively, the security measure includes generating a vehicle guidance signal for intervening in longitudinal guidance and/or lateral guidance of a motor vehicle for an automated driving function. This makes it possible to intervene automatically in the driving operation of a motor vehicle in order, for example, to trigger an evasive maneuver (steering through longitudinal guidance) and/or emergency braking (steering through lateral guidance) and/or emergency braking (deacceleration or braking through longitudinal guidance). For this purpose, the processor circuit can be operated in the motor vehicle. An autonomous driving function (so-called Level 3 and higher according to the SAE J3016 standard) can be supported as an automated driving function.

As a further solution, the invention includes a processor circuit for a motor vehicle, the processor circuit being adapted to carry out a method according to one of the preceding claims. The processor circuit can be designed, for example, as a control device or as a network of several control devices or as a central computer for a motor vehicle. Alternatively, the processor circuit can be designed as part of a stationary backend server for a motor vehicle. The processor circuit can also be implemented as a distributed circuit partly in a backend server and partly in a motor vehicle. The processor circuit can represent a data processing device or a processor device. It can be set up to carry out an embodiment of the method according to the invention. The processor circuit can do at least one Microprocessor and / or at least one microcontroller and / or at least one FPGA (Field Programmable Gate Array) and / or at least one DSP (Digital Signal Processor). Furthermore, the processor circuit can have program code that is designed to carry out the embodiment of the method according to the invention when executed by the processor circuit. The program code can be stored in a data memory of the processor circuit.

As a further solution, the invention includes a motor vehicle with an automated driving function and with at least one environmental sensor for generating sensor data from objects in an environment of the motor vehicle and with an image analysis circuit for generating input data from the sensor data, the processor circuit described for controlling the automated driving function is provided, which is coupled to the image analysis circuit. The image analysis circuit can be used in a manner known per se to extract the information from sensor data or image data from the environmental sensor where an object is represented in an image of the environment, for example a camera image. The image analysis circuit can then define, for example, a bounding box as input data for the downstream artificial neural network, which indicates which sensor data, for example which image data, presumably represent a single object that is detected by the artificial neural network, that is, by its object recognition , should be detected or recognized by specifying the associated object class as the detection result. Corresponding image analysis circuits are known per se from the prior art and can be adopted from the prior art. The image analysis circuit can also be implemented as a component of the processor circuit, that is, the image analysis for generating the input data and the downstream artificial neural network for object recognition can be formed by the same processor circuit and, for example, implemented as an entire or common artificial neural network. The motor vehicle according to the invention is preferably designed as a motor vehicle, in particular as a passenger car or truck, or as a passenger bus or motorcycle.

The invention also includes the combinations of the features of the described embodiments. The invention therefore also includes implementations that each have a combination of the features of several of the described embodiments, provided that the embodiments have not been described as mutually exclusive.

• 1 eine schematische Darstellung einer Ausführungsform des erfindungsgemäßen Kraftfahrzeugs;
• 2 ein Flussschaudiagramm zur Veranschaulichung einer Ausführungsform des erfindungsgemäßen Verfahrens;
• 3 ein Flussschaudiagramm zur Veranschaulichung weiterer Aspekte des Verfahrens.

Exemplary embodiments of the invention are described below. This shows:

• 1 a schematic representation of an embodiment of the motor vehicle according to the invention;
• 2 a flowchart to illustrate an embodiment of the method according to the invention;
• 3 a flowchart to illustrate further aspects of the process.

The exemplary embodiments explained below are preferred exemplary embodiments of the invention. In the exemplary embodiments, the components described each represent individual features of the invention that can be viewed independently of one another, which also develop the invention independently of one another and are therefore to be viewed as part of the invention individually or in a combination other than that shown. Furthermore, the exemplary embodiments described can also be supplemented by further features of the invention that have already been described.

In the figures, functionally identical elements are each provided with the same reference numerals.

1 shows a motor vehicle 10, which can be a motor vehicle, in particular a passenger car or truck. The motor vehicle 10 can carry out a journey 11 through an environment 12 and can include a respective current traffic situation 13 by means of at least one environmental sensor 14, the detection area 14 'of which can be aligned with the environment 12. The environmental sensor 14 can be provided as a camera, for example. The respective environmental sensor 14 can generate sensor data 15 through which the environment 12 is mapped. Objects 16 can be located in the environment 12, for example other road users and/or traffic infrastructure, such as traffic lights.

An automated driving function 17 can be provided in the motor vehicle 10, through which at least part of the leadership task or the complete leadership task for guiding the motor vehicle 10 in the environment 12 (longitudinal guidance and/or lateral guidance) can be carried out during the journey 11. The automated driving function 17 can be carried out, for example, by one control device or several control devices or a central computer of the motor vehicle 10 become. In order to drive the motor vehicle 10, the automated driving function 17 can be provided with detection data 18 by a processor circuit 19, which can detect or recognize from the sensor data 15 which objects 16 are in the environment 12. Through the detection data 18, a corresponding object class 20, for example the indication that it is another motor vehicle or a traffic light, and for example a relative position (not explained in more detail here) can be signaled to the automated driving function 17. The automated driving function 17 can, for example, plan a travel trajectory 21 based on the detection data 18 and drive the travel trajectory 21 in the environment 12 in a manner known per se by controlling at least one actuator 22 of the motor vehicle 10.

For example, an image analysis circuit 23 can be implemented within the processor circuit 19, which can detect which portion of the sensor data 15, for example which image section from a camera image, represents or images one of the objects 16. For each individual object 16, corresponding input data 24 can then be extracted from the sensor data 15 and provided to an artificial neural network 25, which determines the detection data 18, that is, the information about which object class it is, as a detection result can.

The artificial neural network can be provided, for example, by a backend server 26, which can be operated, for example, on the Internet 27. The backend server 26 can, for example, be coupled via a radio network 28, for example a mobile radio network and/or a WLAN radio network (WLAN - Wireless Local Area Network), via a radio connection 29 to a communication module 30 of the motor vehicle 10, via which the processor circuit 19 communicates data with the Backend server 26 can exchange. A corresponding communication connection can be based, for example, on an Internet connection. The communication module 30 can include, for example, a mobile radio module or a WLAN radio module.

The processor circuit 19 ensures that the detected object class 20 has been checked again, that is to say a detection result D of the artificial neural network 25 can also have been checked again.

2 schematically illustrates a corresponding procedure. The environmental sensor 14 can use the sensor data 15 in the manner described the environment 12, for example a camera image. The environmental sensor 14 can be designed accordingly as a camera, in particular as a video camera, that is to say the environment 12 can be described in a sequence or successive individual images using the sensor data 15. By means of the image analysis circuit 23, individual objects 16, which are in the are shown, a delimitation or segmentation takes place, so that each object 16 is identified, for example, by a bounding box 32 or in the can be delimited. The sensor data 15 determined for the respective object 16, for example the respective image section in the respective bounding box 32, can represent input data 24 that can be provided to the network 25. In the network 25, layers 34 of artificial neurons of the neural network 25 can be divided in such a way that a feature extractor 35 and a detector 36 are formed in a manner known per se. The feature extractor 35 can be implemented on the basis of a convolutional network CNN, the detector 36 on the basis of fully networked layers FC (Fully Connected) in a manner known per se.

Between the feature extractor 35 and the detector 36 there can be a so-called bottleneck or bottleneck 37, which is how the interface between a feature extractor and a detector is called. The layer present in the area of the bottleneck 37 or the layer representing the output of the feature extractor 35 can represent a predetermined set 38 of activation values of this layer, which can be transferred to the detector 36 for object recognition.

To verify or check the plausibility of the resulting detection result D, the set 38 of the activation values can be additionally checked. For this purpose, a series of numbers or a signal 41 can be generated from the set 38 of activation values by unfolding 40 or concatenation, from which a feature vector 43 can be generated by means of a linear model 42 for multivariable statistics, in particular by means of a principal component analysis PCA or PC1, which can, for example, provide dimensions X1, X2.

It can now be checked whether the feature vector 43 is typical or probable for the prediction or the detection result D.

For this purpose, depending on the corresponding detection data D ', which indicates the detection result D, a statistical frequency model 45 assigned to the detection result D from a data memory 44 can be created from several, for which different object classes 20 provided frequency models can be selected. The frequency model 45 can, for example, be formed from several kernels 46 of Gaussian distributions or it can be a support vector machine.

For the frequency model 45, which corresponds to the object class of the detection result D, the feature vector 43 can now be checked to see which class probability information 47 results for the feature vector 43 in relation to the frequency model 45 (eg as a probability value or probability density value X3). The class probability information 47 can then be converted into the control signal 48 or mapped onto a signal value of the control signal 48. By means of the control signal 48, a security measure 49 can be controlled, which ensures or contributes to the fact that the detection data 18 ( 1 ), which are transferred to the automated driving function 17, are plausible or reliable.

3 illustrates once again the check by which a class probability information 47 results from the feature vector 43.

The idea is thus to introduce a reliability metric to find the anomalous data points based on their distance from the center of mass of the clean data of the class suggested by the network, preferably a DNN. Based on this, we invent an in-vehicle system that uses a reference cluster model as a frequency model built in the backend based on the training data set used to optimize the DNN to check whether the current prediction of the DNN is reliable or not not. At each stage of perception and decision-making in which DNN makes predictions, the reliability of these predictions is crucial for the entire automated driving system, as the system's upcoming decisions can be influenced by these reliability values. In other words: If a prediction of a subsystem, e.g. B. perception, proves to be unreliable, the system must make alternative decisions instead of taking such predictions into account, otherwise the safety of passengers or other road users may be endangered.

The state of the art is based on computationally intensive methods, which are difficult given the limited resources of the operational devices, such as. B. vehicles, are not always easy to use. On the other hand, their predictions are often misleading during unfavorable perturbations as they still provide a high level of reliability for the incorrectly predicted data points. Our method is based on lightweight statistical models that require only a fraction of the computing power required by the main DNN and thus do not affect the efficiency of the entire detection system. Furthermore, because it is based on statistical analysis methods, system engineers can set decision boundaries to rely only on a certain range of valid detections and consider the rest as unreliable. This particularly helps in defining reliable and safe decision ranges within which the DNN can perform well.

In addition, the method described here can also be used for monitoring perceptual DNNs (detectors), where for the DNNs a variety of input data points with distances to their true class cluster centers (area covered by kernels of a probability function) are evaluated and evidence is generated based on them can be.

• (S1) [Im Backend] werden die Aktivierungen der letzten CNN-Schicht für den gesamten Trainingsdatensatz extrahiert und mit linearen Dimensionalitätsreduktionsmethoden wie der Hauptkomponentenanalyse (PCA) vereinfacht, was einen vereinfachten Merkmalsraum ergibt, der für verschiedene Klassen geclustert werden kann. Aufgrund seiner Linearität ist dieses Modell beträchtlich klein und leicht, was eine minimale Berechnungsleistung erfordert.
• (S2) [Im Backend] Auf der Grundlage der Ergebnisse aus (S1) wird ein Gauß-basiertes Clustermodell (Gauß-kernel-basierte Verteilungsfunktion) entwickelt, das für jede Klasse ein Cluster bildet. Dieses wird dann verwendet, um einen Wahrscheinlichkeitsscore zu schätzen, der die Wahrscheinlichkeit definiert, dass ein Datenpunkt zu einem Cluster gehört. Dieser Wahrscheinlichkeitsscore wird so berechnet, dass die Rechenleistung des Fahrzeugs möglichst gering ist.
• (S3) [Im Backend] wird für jedes Cluster ein Wahrscheinlichkeitsschwellenwert β definiert, der die Mindestwahrscheinlichkeitsbewertung darstellt, die eine positive Klassenvorhersage haben kann.
• (S4) [Im Fahrzeug] werden für jeden neuen Datenpunkt, der dem DNN zugewiesen wird, dieselben CNN-Schichtaktivierungen extrahiert und dem Dimensionalitätsreduktionsmodell aus (S1) zugewiesen.
• (S5) [Im Fahrzeug] wird die Wahrscheinlichkeitsbewertung der Ergebnisse aus (S4) in Bezug auf das Klassencluster unter Verwendung der vorgeschlagenen Klassen-ID aus der endgültigen Vorhersage des DNNs geschätzt.

• (6) [Im Fahrzeug] Eine kann nun den Signalwert für das Steuersignal 49 erzeugen: Wenn der in Schritt (S5) geschätzte Wahrscheinlichkeitswert unter dem in Schritt (S3) berechneten Schwellenwert liegt, wird die endgültige Vorhersage als „nicht zuverlässig“ betrachtet.

An overview of the preferred method is in 3 shown. The decision as to whether a prediction made by the automated driving system in the vehicle is reliable or not can be made using the following six steps S1 to S6. Steps S1 to S3 are preferably carried out during the design time of the system in the backend, while steps S4 to S6 are carried out during runtime in the motor vehicle, which requires a minimum of computing power. The steps can be designed as follows:

• (S1) [In the backend], the activations of the final CNN layer for the entire training data set are extracted and simplified using linear dimensionality reduction methods such as principal component analysis (PCA), resulting in a simplified feature space that can be clustered for different classes. Due to its linearity, this model is considerably small and lightweight, requiring minimal computational power.
• (S2) [In the backend] Based on the results from (S1), a Gaussian-based cluster model (Gaussian kernel-based distribution function) is developed, which forms a cluster for each class. This is then used to estimate a probability score, which defines the probability that a data point belongs to a cluster. This probability score is calculated so that the computing power of the vehicle is as low as possible.
• (S3) [In the backend] a probability threshold β is defined for each cluster, which represents the minimum probability score that a positive class prediction can have.
• (S4) [In Vehicle] For each new data point assigned to the DNN, the same CNN layer activations are extracted and assigned to the dimensionality reduction model from (S1).
• (S5) [In Vehicle], the probability score of the results from (S4) with respect to the class cluster is estimated using the proposed class ID from the final prediction of the DNN.

• (6) [In vehicle] One can now generate the signal value for the control signal 49: If the probability value estimated in step (S5) is below the threshold value calculated in step (S3), the final prediction is considered “not reliable”.

The method is applicable to various recognition tasks such as semantic segmentation, 2D and 3D object detection. The results can be used for security improvement of DNNs based on data points of different difficulty levels.

Reference symbol list

10

motor vehicle

11

Drive

12

Vicinity

13

Traffic situation

14

Ambient sensor

15

Sensor data

16

object

17

Driving function

18

Detection data

19

Processor circuit

20

Object class

21

Travel trajectory

22

actuator

23

Image analysis circuit

24

Input data

25

network

26

Backend server

27

Internet

28

Radio network

29

Radio connection

30

Communication module

31

Illustration

32

Bouding box

34

layers

35

Feature extractor

36

detector

37

bottleneck

38

Crowd

40

unfolding

41

signal

42

Linear model

43

Feature vector

44

Data storage

45

Frequency model

46

kernel

47

Class probability statement

48

Control signal

49

Hedging measure

D

Detection result

QUOTES INCLUDED IN THE DESCRIPTION

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

Cited patent literature

• DE 19720764 A1 [0009]
• DE 102006003611 A1 [0010]
• DE 102018211099 A1 [0011]
• DE 102019116801 A1 [0012]

Claims (10)

Method for checking the plausibility of a detection result of an object recognition in an artificial neural network (25), wherein the network (25) receives input data (24) relating to an object (16) that is detected by sensors in an environment (12). , • as a function of the input data (24) in several layers (34) of the network (25) activation values of artificial neurons of the network (25) are determined and • as a function of the activation values at an output layer of the network (25) as a detection result, detection data (18 ) are generated, which indicate which object class (20) from several predetermined object classes (20) detectable by the network (25) was detected for the object (16), characterized in that by a processor circuit (19) • in addition to the Detection data (18) from a predetermined set (38) of the activation values of at least one of the layers (34) a feature vector (43) is formed and • a statistical frequency model (45) assigned to the detected object class (20) is selected and • using the frequency model ( 45) a class probability information (47) is determined, which indicates how likely the feature vector (43) is for the detected object class (20) according to the selected frequency model (45), and • the class probability information (47) using a predetermined mapping function on a control signal (48) (41) is depicted, which a predetermined security measure (49) is controlled to compensate for false detection. Procedure according to Claim 1 , wherein the set (38) of activation values is taken from one layer or several layers (34), each of which has an output of a feature extractor (35) of the network (25), to which a detector (36) of the network (25) is connected downstream . Method according to one of the preceding claims, wherein a dimension of the feature vector (43) is smaller than a number of activation values in the set (38) of activation values and for this purpose the set (38) of activation values is calculated using a linear model (42) for multivariable statistics the feature vector (43) is mapped. Procedure according to Claim 3 , where principal component analysis, PCA, is used as the linear model (42) of multivariable statistics. Method according to one of the preceding claims, wherein the feature vector (43) has dimensionality 3. Method according to one of the preceding claims, wherein the respective frequency model (45) • comprises a kernel-based distribution function of a frequency distribution, in particular a Gaussian kernel-based distribution function, and a probability value or probability density value is determined as class probability information (47) and/or • a support vector machine for a classification analysis of the feature vector (43) and a binary membership statement is determined as the class probability information (47). Method according to one of the preceding claims, wherein the mapping function is used to check whether the feature vector (43) confirms the object (16) with a predetermined minimum probability according to the class probability information (47). Method according to one of the preceding claims, wherein the security measure (49) comprises that • a spatial resolution and/or a time resolution for at least one environmental sensor (14), which generates the input data (24), is enlarged and/or • Predetermined computing resources are switched on to generate the input data (24) and/or • a vehicle guidance signal (41) is generated to intervene in a longitudinal guidance and/or lateral guidance of a motor vehicle (10) for an automated driving function (17). Processor circuit (19) for a motor vehicle (10), the processor circuit (19) being adapted to carry out a method according to one of the preceding claims. Motor vehicle (10) with an automated driving function (17) and with at least one environmental sensor (14) for generating sensor data (15) of objects (16) in an environment (12) of the motor vehicle (10) and with an image analysis circuit (23). Generating input data (24) from the sensor data (15), wherein a processor circuit (19) is provided for controlling the automated driving function (17), which is coupled to the image analysis circuit (23) and according to Claim 9 is designed.

Images