Granted Patent B2

US 12,711,748 · App. 18/046,087 · Granted Aug 18, 2026

Inventors: Gesina Schwalbe (Regensburg, DE); Christian Wirth (Niederhöchstadt, DE)

Assignee: AUMOVIO Germany GmbH

G06V10/776G06V10/82

Abstract

Claims (40)

1 . A computer-implemented method for monitoring a logical consistency of an artificial neural network which is configured for a classification, classification and localization, object detection, or segmentation of features which are contained in input data, and for outputting final output data which are indicative of the existence and, if applicable, a position of the features in the input data, the method comprising:

(a) reading in activation data of the artificial neural network, which are produced from the input data, by an input interface;

(b) transferring the activation data read in (a) to at least one trained concept model which is trained to recognize and, if applicable, localize a partial feature of the features contained in the input data, and to output a partial feature mask, wherein the partial feature mask contains at least one partial feature truth value which is indicative of whether and, if applicable, where the partial feature is contained in the input data;

(c) linking the final output data to the partial feature truth value by a fuzzy logic unit which determines at least one continual logical consistency truth value therefrom;

(d) evaluating the at least one continual logical consistency truth value obtained in (c) by an evaluation unit, wherein, if the at least one continual logical consistency truth value reaches or falls short of a predefined threshold, the evaluation unit ascertains a logical inconsistency of the final output data in an inconsistency region, wherein the evaluation unit ascertains a logical consistency if the at least one continual logical consistency truth value does not fall short of the predefined threshold; and

(e) in the event that no logical consistency of the final output data with the input data was ascertained in (d), activating additional sensors or controlling a vehicle component.

2 . The method according to claim 1 , further comprising in (e) performing, at least one additional action of the following:

outputting an uncertainty measure, optionally localized in the inconsistency region;

performing a redundant evaluation of the input data;

activating further control mechanisms for the artificial neural network;

adopting a safe state; or

displaying a consistency error.

3 . The method according to claim 1 , wherein in (a), the activation data are read in from an output layer and/or at least one intermediate layer of the artificial neural network.

4 . The method according to claim 1 , wherein in (b), the activation data are transferred from one layer of the artificial neural network to the at least one trained concept model.

5 . The method according to claim 1 , wherein in (b), the at least one trained concept model has been trained by a linear machine learning method as a linear machine learning model.

6 . The method according to claim 1 , wherein

in (b), the activation data are transferred to a plurality of trained concept models, wherein the partial feature mask of each trained concept model is linked to one another in order to obtain a total partial feature mask with total partial feature truth values, wherein

in (c), the final output data are linked to the total partial feature truth values by the fuzzy logic unit.

7 . The method according to claim 6 , wherein in (c), the at least one continual logical consistency truth value is determined by applying an implication link to one of the partial feature mask or the total partial feature mask and the final output data, with a proviso that the one of the partial feature mask or the total partial feature mask implies at least regions of the final output data.

8 . The method according to claim 1 , wherein the input data contain input image data, wherein the artificial neural network is configured for a classification, classification and localization, object detection, or segmentation of the input image data into feature regions, wherein the feature regions contain recognized features which characterize a semantic concept, wherein the final output data contain the segmented feature regions, and wherein

in (a), the activation data, in the form of activation values or activation maps of the artificial neural network, which are produced in each case from the input image data, are read in by the input interface;

in (b), the at least one trained concept model is trained to recognize and, if applicable, localize a partial feature region of the feature regions contained in the input image data and to output the partial feature mask, wherein the partial feature mask contains at least one region of the at least one continual partial feature truth value which are indicative of whether and where a partial feature range is contained in the input image data;

in (c), the final output data are linked pixel-by-pixel to the at least one continual partial feature truth value by the fuzzy logic unit which determines a consistency truth value map of the at least one continual logical consistency truth value therefrom; and

in (d), the consistency truth value map obtained in (c) is evaluated by the evaluation unit, wherein, in a region of the consistency truth value map in which the at least one continual logical consistency truth value reaches or falls short of a predefined threshold, the evaluation unit ascertains a logical inconsistency of the final output data with the input image data within the inconsistency region, wherein the evaluation unit ascertains a logical consistency within a region of the consistency truth value map if the at least one continual logical consistency truth value in this region do not fall short of the predefined threshold.

9 . The method according to claim 8 , wherein the input image data contain an environment situation of a motor vehicle, in particular a traffic situation with at least one person, the artificial neural network is configured for a classification, classification and localization, object detection, or segmentation of the input image data into feature regions which contain persons, traffic signs and/or road markings, the final output data contain the segmented feature regions, and wherein

in (b), a plurality of trained concept models are trained to recognize and, if applicable, localize partial features of the persons, traffic signs and/or road markings of the feature regions contained in the input image data, and to output the partial feature mask.

10 . The method according to claim 9 , wherein at least one additional sensor or one additional camera is activated, which is configured to detect persons, traffic signs and/or road markings; and/or input image data are detected again and/or evaluated again, by another method.

11 . A computer-implemented method for recognizing and, if applicable, localizing objects and/or persons in a traffic situation,

wherein input data are processed by a machine learning model which is trained for a classification, classification and localization, object detection, or segmentation of the input data, in order to obtain final output data;

wherein a method according to claim 1 is performed, in parallel or a temporally subordinate manner, by a monitoring device in order to obtain a consistency truth value map; and

wherein a redundant, supplementary and/or renewed detection and/or evaluation of the input data is/are carried out as a function of the consistency truth value map.

12 . A monitoring device for monitoring the logical consistency of a machine learning model, wherein the monitoring device comprises components including the input interface, the at least one trained concept model, the fuzzy logic unit and the evaluation unit, wherein the monitoring device and the components thereof are configured to perform a method according to claim 1 .

13 . A system for recognizing and, if applicable, localizing objects and/or persons in a traffic situation, wherein the system comprises a sensor device configured for detecting the input data, the machine learning model configured for recognizing objects and/or persons and the monitoring device according to claim 12 for monitoring the machine learning model.

14 . A motor vehicle having a system according to claim 13 .

15 . A non-transitory computer-readable storage medium comprising commands which, when executed by a computer;

(a) read in activation data of an artificial neural network via an input interface, the artificial neural network being configured for a classification, classification and localization, object detection, or segmentation of features which are contained in input data, and for outputting final output data which are indicative of the existence and, if applicable, a position of the features in the input data, the activation data being produced from the input data;

(b) transfer the activation data read in (a) to at least one trained concept model which is trained to recognize and, if applicable, localize a partial feature of the features contained in the input data, and to output a partial feature mask, wherein the partial feature mask contains at least one partial feature truth value which is indicative of whether and, if applicable, where the partial feature is contained in the input data;

(c) link the final output data to the partial feature truth value by a fuzzy logic unit which determines at least one continual logical consistency truth value therefrom;

(d) evaluate the at least one continual logical consistency truth value obtained in (c) by an evaluation unit, wherein, if the at least one continual logical consistency truth value reaches or falls short of a predefined threshold, the evaluation unit ascertains a logical inconsistency of the final output data in an inconsistency region, wherein the evaluation unit ascertains a logical consistency if the at least one continual logical consistency truth value does not fall short of the predefined threshold; and

(e) in the event that no logical consistency of the final output data with the input data was ascertained in (d), activating additional sensors or controlling a vehicle component.

Assignments (3)

Priority Claims (1)

Continuity (1)

References Cited (34)