Detecting and determining relevant variables of an object by means of ultrasonic sensors
The present disclosure relates to a method, to a computer program comprising instructions, and to a device for environment sensing in a vehicle. For the environment sensing, environment data are recorded by means of at least one vehicle sensor. A prediction is calculated by means of a trained machine learning model based on the recorded environment data, wherein the prediction includes a measure of uncertainty for the prediction. Equally, a deviation value is determined for the recorded environment data that provides a measure of how significantly the recorded environment data deviate from training data for the machine learning model. At least one conformity score is determined based on the measure of uncertainty and the deviation value and then a prediction set is determined based on the at least one determined conformity score. Then, a control signal is generated depending on the determined prediction set.
1 . A method for environment sensing in a vehicle, comprising:
recording environment data using at least one vehicle sensor;
calculating a prediction using a trained machine learning model based on the recorded environment data, wherein the prediction includes a naive measure of confidence, which can be based on one or more of an activation function for an outputting neurons in an output layer of a neural network, and an error predictor;
determining a deviation value for the recorded environment data that provides a measure of how significantly the recorded environment data deviate from training data for the machine learning model;
determining at least one conformity score based on the measure of uncertainty and the deviation value;
determining a prediction set based on the at least one determined conformity score and previously determined calibration parameters; and
generating a control signal for a vehicle function depending on the determined prediction set.
2 . The method of claim 1 , wherein, after the prediction set has been determined, a distinction is made as to whether the prediction set is empty or contains at least one predictor, wherein different control signals are generated based on the distinction.
3 . The method of claim 2 , wherein a control signal is generated in the case that the prediction set is empty in order to initiate an emergency maneuver for the vehicle.
4 . The method of claim 2 , wherein a control signal is generated in the case that the prediction set comprises at least one predictor in order to prompt the vehicle to react according to the at least one predictor.
5 . The method of claim 4 , wherein a reaction takes place in such a manner permitted by the highest risk class for the predictors contained in the current prediction set.
6 . The method of claim 1 , wherein the at least one conformity score is determined based on a function that provides higher conformity scores for higher deviation values and/or higher values for the measure of uncertainty.
7 . The method of claim 6 , wherein the deviation values and the values for the measure of uncertainty are additively linked in the function.
8 . The method of claim 6 , wherein the deviation values and the values for the measure of uncertainty are multiplicatively linked in the function.
9 . The method of claim 6 , wherein the trained machine learning model is configured for environment sensing as a neural network, wherein a measure of confidence in the predictions of the neural network results from the activation function for outputting the neurons in the output layer of the neural network and wherein the measure of uncertainty results from 1−measure of confidence.
10 . The method of claim 1 , wherein the method is carried out in real time for environment sensing in a vehicle.
11 . The method of claim 1 , wherein the calibration parameters are determined during a calibration by means of the distribution of the conformity scores.
12 . A computer program comprising instructions which, when executed by a computer, cause the computer to:
record environment data using at least one vehicle sensor;
calculate a prediction using a trained machine learning model based on the recorded environment data, wherein the prediction includes a naive measure of confidence, which can be based on one or more of: an activation function for outputting neurons in an output layer of a neural network, and an error predictor;
determine a deviation value for the recorded environment data that provides a measure of how significantly the recorded environment data deviate from training data for the machine learning model;
determine at least one conformity score based on the measure of uncertainty and the deviation value;
determine a prediction set based on the at least one determined conformity score and previously determined calibration parameters; and
generate a control signal for a vehicle function depending on the determined prediction set.
13 . A device for environment sensing in a vehicle, comprising
a recording circuit for recording environment data using at least one vehicle sensor; and
a sensing circuit that is configured to:
calculate a prediction using a trained machine learning model based on the environment data recorded using the recording circuit, wherein the prediction includes a naive measure of confidence, which can be based on one or more of: an activation function for outputting neurons in an output layer of a neural network, and an error predictor;
determine a deviation value for the recorded environment data that provides a measure of how significantly the recorded environment data deviate from training data for the machine learning model;
determine at least one conformity score based on the measure of uncertainty and the deviation value;
determine a prediction set based on the at least one determined conformity score and previously determined calibration parameters; and
generate a control signal for a vehicle function depending on the determined prediction set.
14 . The device of claim 13 , wherein the sensing circuit is implemented on a controller of the vehicle.
15 . A vehicle which comprises the device of claim 13 or is configured to carry out the method of claim 1 .
16 . The method of claim 3 , wherein a control signal is generated in the case that the prediction set comprises at least one predictor in order to prompt the vehicle to react according to the at least one predictor.
17 . The method of claim 16 , wherein a reaction takes place in such a manner permitted by the highest risk class for the predictors contained in the current prediction set.
18 . The method of claim 2 , wherein the at least one conformity score is determined based on a function that provides higher conformity scores for higher deviation values and/or higher values for the measure of uncertainty.
19 . The method of claim 3 , wherein the at least one conformity score is determined based on a function that provides higher conformity scores for higher deviation values and/or higher values for the measure of uncertainty.
20 . The method of claim 4 , wherein the at least one conformity score is determined based on a function that provides higher conformity scores for higher deviation values and/or higher values for the measure of uncertainty.